Entries get the research, exits get a guess

The standard pattern in both retail and academic backtests is a carefully engineered entry paired with a symmetric \(+X\% / -Y\%\) bracket or a fixed holding period, with \(X\), \(Y\) or the horizon picked by grid search over the same sample used to judge the entry. Three things go wrong at once.

The exit grid is a multiple-testing problem nobody reports. A 10 × 10 grid of take-profit and stop-loss levels is 100 trials. Under the null, the expected maximum of \(N\) roughly independent Sharpe estimates is about \(\sigma_{\widehat{SR}}\,\big[(1-\gamma)\,\Phi^{-1}(1 - 1/N) + \gamma\,\Phi^{-1}(1 - 1/(Ne))\big]\), where \(\gamma \approx 0.577\) is the Euler–Mascheroni constant. That is about \(2.5\,\sigma_{\widehat{SR}}\) for \(N = 100\). Five years of daily data give \(\sigma_{\widehat{SR}} \approx 1/\sqrt{5} \approx 0.45\) annualized, so the exit grid alone can manufacture an annualized Sharpe near 1.1 from a signal with no edge. Neighbouring grid cells are correlated, so the effective \(N\) is smaller, but rarely small enough to ignore. Put the full grid size into the deflated Sharpe calculator, not just the number of entry variants. The deflated Sharpe guide covers the derivation.

The optimum sits on noise. A sharp peak in the stop-multiplier heatmap is the signature of fitting noise. A real exit effect looks like a broad plateau. The neighbourhood test applies to exits with even more force than to entries, because exit parameters interact with path detail the entry never sees.

Brackets on bars carry a look-ahead bug. When both the take-profit and the stop fall inside one bar’s high–low range, the bar does not say which was hit first. Many backtesters resolve this toward the take-profit, which is same-bar look-ahead concentrated in volatile bars, where tight brackets get tested most. The conservative rule is to assume the stop filled first, or to resolve the order from tick data. Tight brackets on bar data should be treated as suspect until proven otherwise.

Fixed time exits have their own version of the problem. “Exit after \(n\) bars” with \(n\) tuned in-sample is the same search, and it also hard-codes the assumption that the signal’s information decays on the calendar instead of with the market’s activity clock (see information-driven bars).

Exits as thesis invalidation: three structural anchors

The alternative is to define exits by what would have to be true for the entry’s reason to stop holding, and to choose anchors whose triggers respond to different features of the market. Take an intraday trend entry: a breakout to a new session high, beyond a volatility band anchored at the open. There are three natural ways to say “this is no longer a trend trade.”

1. Volume-weighted anchor: session or rolling VWAP.

VWAP is the average price paid by the session’s participants, weighted by how much they traded. Price returning to it means the imbalance that drove the breakout has been absorbed, and the market again agrees on the volume-weighted fair value. The anchor follows where volume printed, so a heavy-volume morning pulls it toward the morning’s prices.

2. Range-midpoint anchor: session midline or Donchian midpoint.

The midline ignores volume entirely. It asks whether price is still nearer the session’s high-water mark than its low. For a long entered at a new high, \(M_t\) ratchets up with every new high, so the anchor is an implicit trailing stop at half the range. It catches structural failures that VWAP misses on days when volume was concentrated far from the current price.

3. Time-accelerating trailing stop: Parabolic SAR or an explicit convex time hurdle. Wilder’s SAR updates

where EP is the extreme price since entry, and the acceleration factor AF steps up by 0.02 at each new extreme. The stop tightens faster the longer and further the trade runs. The general principle is an exit hurdle that rises with holding time. A cleaner, explicit version for a long is

which gives the trade room early and, as \(t \to \tau_{\max}\), demands at least \(\theta\) of profit. The rationale is economic. A breakout signal’s expected remaining drift decays with time, while a stagnant position keeps its full variance, its share of the risk budget, and its exposure to whoever is on the other side. A trade that has gone nowhere for an hour has a lower ex-ante Sharpe than one that just triggered, and the hurdle prices that in.

The three anchors fire on different evidence: where volume traded, the geometry of the range, and elapsed time. Their exit times are correlated, because a sharp reversal crosses all three, but they are not redundant. That partial independence is what makes combining them worthwhile.

There is published evidence for the rotation as well. A 2026 study by Concretum Group of an open-anchored ATR-band breakout on a semiconductor ETF (2012–2026) tested four exits: session open, midline, VWAP and PSAR. Every variant was profitable, with full-sample Sharpe ratios of 0.92 to 1.16. Annual rankings rotated with no persistent winner, and per-trade statistics traded off against each other. The exit with the highest average gross return per trade (about 6.8 bps) also had the deepest average intra-trade drawdown (about −89 bps), against roughly 6.1 bps and −64 bps for VWAP. Fourteen years of data put the standard error of an annualized Sharpe at about \(1/\sqrt{14} \approx 0.27\), so the whole spread between the best and worst exit is less than one standard error. Picking the top one would have been picking noise.

Fractional ensembling vs winner-take-all: the arithmetic

Fix the entry. Let \(R_k\) be the trade’s return under exit rule \(k\), for \(k = 1, \dots, K\). The fractional ensemble liquidates \(1/K\) of the position when each rule triggers, so

assuming roughly equal variances \(\bar\sigma^2\) and average pairwise correlation \(\bar\rho\).

Variance. With \(K = 3\) and \(\bar\rho = 0.7\), the variance is 80% of a single rule’s, so the volatility is about 11% lower. The mean is the average of the members’ means. At equal Sharpe across members, the ensemble’s Sharpe is higher by \(1/\sqrt{\bar\rho + (1-\bar\rho)/K} \approx 1.12\times\). An ensemble of exits averaging Sharpe 1.05 therefore lands near the best single member without anyone having to know which member that is.

Selection. Winner-take-all picks \(k^{*} = \arg\max_k \hat\mu_k\), where each in-sample estimate \(\hat\mu_k = \mu_k + e_k\) carries noise \(e_k\) with standard error \(s\). When the true differences between the \(\mu_k\) are small relative to \(s\), as in the published study above, the choice is close to random. The expected out-of-sample return of the pick is then about the members’ average, the same as the ensemble. The in-sample estimate is still inflated by \(\mathbb{E}[\max_k e_k]\), which is about \(1.03\,s\) for \(K = 4\). Winner-take-all delivers the same expected return as the ensemble, with more variance and a reported number that is biased upward. The ensemble involves no selection, so its in-sample estimate is not biased in this way. This is the same reason forecast combination and strategy ensembling keep beating “pick the best” in honest evaluations: the uncertainty about which model is best is real, and the allocation should reflect it rather than hide it.

Liquidity cliffs. With square-root impact \(I(Q) \propto \sigma\sqrt{Q/V}\), the cost of exiting quantity \(Q\) in one go scales as \(Q \cdot I(Q) \propto Q^{3/2}\). Split into \(K\) tranches that execute far enough apart for transient impact to decay, the total is \(K\,(Q/K)^{3/2} = Q^{3/2}/\sqrt{K}\). Three tranches cut exit impact by about 42%. The condition matters. When all three anchors trigger in the same bar (a violent reversal through VWAP, midline and SAR together), the ensemble collapses into one full exit. That is the case where a fast exit is correct anyway, and it is also where impact is worst, so size the position for that scenario, not the average one. The capacity estimator and the capacity guide make the size dependence explicit. The square-root impact literature covers why the exponent is close to ½ and how fast the impact decays.

Herding. A single optimized stop tends to land where everyone’s optimized stop lands: round ATR multiples, the prior day’s low, the opening range. Stops clustered at one level form a self-exciting cascade. The first triggers move price into the next ones, and an aggressive exit at that level fills at the worst point of the cascade. Diversified anchors spread your exits across the range.

The execution channel: passive versus aggressive exits

An exit rule says when. The execution channel says at what cost, and for short-horizon strategies the second can matter as much as the first.

Aggressive (taker) exits pay half the spread, the taker fee and slippage. The slippage is conditional: a stop triggers because price is moving against you, often in the middle of a sweep, when the book on your exit side is thin and other stops are firing too. The expected cost is \(c_{\text{taker}} = \tfrac{s}{2} + f_t + \mathbb{E}[\text{slip} \mid \text{trigger}]\), and the last term is larger than its unconditional average for the same reason imbalance-bar closes cluster in thin books.

Passive (maker) exits rest a limit order at or ahead of the anchor. When it fills you earn the half-spread and any rebate. The fills are adversely selected, though: you get filled when price trades through your level, which is the scenario where the exit was needed. When price gaps past the level without trading there, you are not filled at all and have to chase after a larger move. The expected cost is

with \(r\) the rebate, \(m_{\text{adv}}\) the adverse markout after a fill, and \(g\) the average gap past the level when unfilled. Plug in illustrative liquid-ETF numbers: \(s/2 = 0.5\) bp, \(f_t = 0.2\) bp, conditional slippage 1.5 bps, so \(c_{\text{taker}} = 2.2\) bps. For the passive order take \(r = 0.1\), \(m_{\text{adv}} = 1.0\), \(P_{\text{fill}} = 0.8\) and \(g = 6\) bps. Then \(\mathbb{E}[c_{\text{passive}}] = 0.8(0.4) + 0.2(8.2) \approx 2.0\) bps. The means are almost identical, and the distributions are not. The passive exit swaps a reliable 2.2 bps for a mix of small gains and occasional 8+ bp losses, concentrated in the fastest reversals. Passive exits have negatively skewed costs.

The practical answer is to match the channel to the trigger’s urgency. A midline breach on heavy volume is an urgent invalidation and should go out aggressively. A time-hurdle tranche is patient and can rest passively, with a fuse that converts it to a marketable order after a fixed time or price move. Fractional ensembles make this possible, because each tranche can carry its own channel. Passive impact is not zero either, since resting size moves prices too, and it is now being calibrated directly.

The fee wall for intraday exits

The arithmetic is unforgiving at intraday horizons. Take a gross edge of about 6 bps per trade, the level reported in the breakout study above. A breakout entry is aggressive by construction (you are buying a new high, just after the sweep that made it) and costs about 2.2 bps. An aggressive exit costs about 2.2 more. Round trip is about 4.4 bps, which leaves about 1.6 bps net, roughly a quarter of the gross. Small changes now dominate:

• Exit execution 1 bp worse (one more tick of conditional slippage) removes about 60% of the net edge.
• A grid-optimized stop that improves gross per-trade return by 0.5 bp in-sample, which is inside the noise per the selection argument, is wiped out by a 0.5 bp increase in conditional slippage from exiting with the herd.
• Directional hit rate says almost nothing on its own. A 55% win rate with a 1:1 payoff earns \(0.1\) of the average move. Exits change the payoff ratio, the hit rate and the cost per trade together, so evaluate them on net P&L per trade and its distribution, never on win rate.

Run your own turnover and cost assumptions through the transaction-cost break-even calculator. For intraday strategies, exit design and exit execution are where most of the remaining edge gets decided.

A design protocol

1. Fix the entry first. Define exits as invalidations of the entry’s thesis, never as P&L targets derived from the same sample.
2. Choose \(K = 3\)–4 structurally distinct anchors before looking at results, each with a one-sentence economic rationale. If the rationale needs the backtest to justify it, it is a parameter.
3. Allocate equally across anchors. Optimizing ensemble weights reintroduces the selection problem. Equal weights, or inverse tranche volatility, are hard to beat on any sample shorter than a decade.
4. Keep a policy-level catastrophic stop above the ensemble: per-trade risk capped as a fraction of capital (size \(N \le \kappa P / (p_{\text{entry}}\,\delta)\) for risk budget \(\kappa\) and stop distance \(\delta\)), and a daily loss limit across trades. This layer is risk policy, not alpha, and it should be wide enough that it almost never binds.
5. Cost each tranche through its own channel, resolve intrabar order conservatively, and report year-by-year rank stability of the members next to the ensemble.
6. Report the ensemble. If you publish a single best exit, deflate it by the full grid size.

Production failure modes

• Minimum order sizes break the split. One-third of a small position can fall below the venue’s minimum size or lot, so the ensemble silently becomes a single exit. Per-ticket fees punish splitting as well. Check that every tranche is actually tradeable at your smallest live size.
• VWAP definitions drift. Session reset times, whether auction prints are included, and vendor VWAP versus your own computation all move the anchor. Compute it from your own trade feed with the same filters as the backtest.
• Overnight gaps fire everything at once. Every anchor triggers together on a gap, and the conditional cost is the gap itself. Ensembles do not diversify gap risk. Only size and the policy stop address it.
• Stop state is lost on restart. PSAR and convex-hurdle state depend on the path since entry. Persist it, or a restart resets every trailing stop to its loosest level.
• Exchange-native or synthetic stops. Resting exchange stops can be visible or triggered by thin prints. Synthetic stops depend on your uptime during exactly the volatile minutes. Choose deliberately and monitor both.
• Tranche accounting. Three exits per entry means three reconciliation paths. Mismatched partial fills are a common source of phantom positions.

Questions to ask a paper (or backtest)

1. How were exit parameters chosen? Was the grid size included in the multiple-testing correction?
2. With bar data, how is intrabar ordering of stop and target resolved?
3. Are exits costed by channel, with conditional slippage at stops, or with an average cost?
4. Is the reported exit the best of several? Do its year-by-year ranks persist?
5. Does the paper report intra-trade drawdown and holding time, not only the average return per trade?
6. Would the result survive at the paper’s own size given square-root impact on exits, and at 2× costs?

The optimal-stopping literature formalizes some of this. Joint entry and exit timing for mean-reverting spreads is one example, and the drawdown-control-with-restart line connects trade-level exits to strategy-level ones. On the empirical side, a systematic falsification of OHLCV-based intraday signals shows how quickly intraday edges shrink once realistic execution is imposed. The related hubs are statistical arbitrage, HFT & execution, risk management and optimization methods. The strategy-level counterpart to this essay, deciding when to switch the whole strategy off, is the kill-switch essay.

Run your numbers: deflated Sharpe calculator → · transaction-cost break-even → · capacity estimator → · equity-curve simulator → · The foundations: backtest overfitting → · robustness, defined → · transaction costs and impact →

Sources and background: the exit-variant evidence is from Concretum Group’s intraday trend trade management study. The policy, rule and strategy hierarchy for exits follows Trading the Breaking. The point that a signal’s form (feature, forecast, or weights with or without variable net exposure) constrains which exits it can express follows Systematic Long Short. Wilder (1978) for Parabolic SAR. Browse the flowchart library, the guides, and the tools.

Definition. A bar is a summary (open, high, low, close, volume, trade count, notional) of every trade between two closing times \(T_{k-1} < t_i \le T_k\). A sampling rule decides when \(T_k\) arrives. Time bars close on the wall clock. Information-driven bars close once some measure of market activity has accumulated: trades, shares, notional, signed volume, or one-sided aggression. Picking the rule is a modelling decision, and it should match the job the bars are for: volatility estimation, feature engineering, execution scheduling, or event detection.

Why the wall clock is the wrong clock

The cleanest argument is the old subordination idea (Clark 1973; Ané and Geman 2000). Model a time bar’s return as the sum of \(N_k\) trade-level increments, each with variance \(\sigma^2\), where \(N_k\) is the random number of trades that happened to land in that minute. Even if the increments were perfectly Gaussian, the bar return would be a variance mixture:

The excess kurtosis of time-bar returns is therefore at least \(3\,\mathrm{CV}^2(N_k)\), where CV is the coefficient of variation of trades per bar. An intraday U-shape, news bursts and volatility clustering easily push CV above 1 for one-minute bars on liquid names. That gives a kurtosis of 6 or more before any genuinely fat-tailed price move is involved. The same mixture produces heteroskedasticity: a bar’s variance is set by how busy that minute was. Fix \(N_k\) and this channel disappears. That is the whole statistical case for tick bars, and versions of it hold for volume and notional clocks. It is also why realized-variance estimators sampled in intrinsic time can be more efficient than calendar-time sampling, and why time-subordinated Brownian motion remains a live modelling line.

The operational version: 390 one-minute bars give the lunchtime minute with 200 trades the same regression weight as the opening minute with 5,000. A model fit on time bars spends most of its sample on periods where nothing happened. Meanwhile the minutes where order flow was informative (opens, data releases, liquidation cascades) get squeezed into a handful of rows whose closes hide everything that happened inside them.

The sampling rules, formally

Common notation: the trade stream is \(\{(t_i, p_i, v_i)\}\). The aggressor sign \(b_i \in \{-1, +1\}\) comes from the venue’s aggressor flag where one exists. Otherwise it comes from the tick rule (\(b_i = \operatorname{sgn}(\Delta p_i)\), carrying \(b_{i-1}\) forward when \(\Delta p_i = 0\)) or from Lee–Ready against the prevailing quote. A bar closes at the first trade where its accumulator crosses its threshold, and the accumulator then resets.

Tick bars. The accumulator counts trades, \(S_k = \sum 1\), and closes at \(S_k \ge L_{\text{ticks}}\). This equalizes decisions only if one trade is one decision, which is false on most modern venues (see the mechanical channel below).

Volume bars. \(S_k = \sum v_i \ge L_{\text{vol}}\). This equalizes shares or contracts traded, which is the clock that VWAP schedules and the volume-synchronized literature (Easley, López de Prado and O’Hara’s “volume clock”) run on.

Dollar bars. \(S_k = \sum p_i v_i \ge L_{\text{notional}}\). Notional is robust to large price-level changes, splits and share issuance, which matters for multi-year samples. It is the sensible default for crypto, where a coin can move 10× within a sample.

Delta bars. These close when cumulative signed volume crosses a fixed level, \(\lvert\sum b_i v_i\rvert \ge \Delta_0\). They are a fixed-threshold precursor to imbalance bars: simple, interpretable, and blind to whether that level of imbalance is unusual today.

Imbalance bars (TIB / VIB / DIB). Let \(\theta_T = \sum_{t=1}^{T} b_t v_t\), where \(v_t = 1\) gives tick imbalance, \(v_t\) = shares gives volume imbalance and \(v_t = p_t v_t\) gives dollar imbalance. The bar closes when realized imbalance exceeds what is expected for a bar of typical length:

The three expectations are usually exponentially weighted moving averages over past bars and ticks. The rule is a surprise detector: a bar closes early when one side is aggressing harder than the recent base rate.

Run bars. The “run” in López de Prado’s definition is the cumulative one-sided volume inside the bar, not a streak of consecutive same-sign trades:

Here \(\mathbb{P}_0^{\pm}\) is the probability of a buy or sell and \(\bar v^{\pm}\) the expected size conditional on side. Unlike imbalance, a run does not cancel when the other side trades, so run bars respond to sustained one-sided participation (a working meta-order) even while the opposite side keeps printing. Several open-source libraries implement the consecutive-streak version under the same name. Check which one yours computes, because the two produce different bars and different statistics.

Quantile bars. These replace the EWMA expectation with an empirical quantile of a trailing window: close when the accumulator exceeds the \(q\)-th quantile of its value over the last \(W\) completed bars. They are less elegant and more stable, and the trade-off between the window \(W\) and responsiveness is explicit instead of hidden inside an EWMA span.

Two channels: mechanical and informational

A bar series improves for two different reasons, and only one of them carries economic content.

The mechanical channel. Most of the gain from tick, volume and dollar bars is a normalization. They divide out the intraday activity cycle and volatility clustering, so the resulting returns look closer to i.i.d. Gaussian. Part of the measured imbalance is mechanical too. Under the tick rule, bid–ask bounce alone generates alternating signs, so \(\theta_T\) contains spread noise in proportion to how often the price oscillates between bid and ask. On large-tick instruments, depleting the queue at one price level produces a burst of same-sign prints whose “imbalance” mostly reflects queue length. Fragmentation adds another distortion: on most crypto venues a single market order that walks five price levels prints as dozens of trades, one per resting maker order. A tick bar then counts the matching engine’s bookkeeping, not decisions. Before building tick or tick-imbalance bars on such venues, aggregate prints that share an aggressor order ID or an exchange timestamp.

The informational channel. Imbalance and run bars close early when aggression is one-sided beyond its base rate. That is roughly the signature of informed traders splitting meta-orders, liquidations, and toxic flow in general. On this channel the bar close is an event, and the most useful way to treat these bars is as event triggers, not as a resampled dataset. “Sample a label whenever a DIB closes” is a well-posed event study. “Run a GARCH on DIB returns” mixes the trigger into the series being modelled. The informational content they target is the same object studied in the LOB-imbalance literature and the order-flow intensity models, measured with a coarser clock.

Calibration and look-ahead traps

1. The EWMA threshold has an unstable fixed point. Take the imbalance rule under balanced flow: signs close to coin flips and an estimated bias \(\delta = \lvert 2\hat{\mathbb{P}}(b=1) - 1\rvert\) that is small. A symmetric random walk needs about \(h^2\) steps (in units of trade size) to first reach \(\pm h\). With \(h = \mathbb{E}[T]\,\delta\), the next bar’s length is roughly \(T_{k} \approx (\mathbb{E}[T]\,\delta)^2\), and the EWMA feeds it back into \(\mathbb{E}[T]\). The map has a fixed point at \(\mathbb{E}[T]^{*} = 1/\delta^{2}\), and it is unstable. Start above it and bars lengthen without bound. Start below it and they collapse toward one-trade bars. Which way a backtest goes depends on the warm-up guess, and that is the “bars explode on day three” behaviour practitioners report. The analysis is stylized (\(\delta\) is itself re-estimated), but the instability is real. Fixes: anchor \(\mathbb{E}[T]\) to a fixed budget and adapt only \(\delta\) and \(\mathbb{E}[v]\); clamp \(\mathbb{E}[T]\) to \([T_{\min}, T_{\max}]\); or switch to quantile thresholds.

2. Halts and auctions break volume and notional clocks. An opening or closing auction prints one trade carrying a large share of the day’s volume. A dollar-bar threshold set as a fraction of daily notional is blown through by a single print, which yields either one enormous bar or many zero-duration bars sharing one timestamp, depending on whether the implementation splits prints. After a volatility halt the EWMA still holds the pre-halt regime, so the first post-halt bars are mis-sized in whichever direction the halt moved activity. Route auction prints to their own bar type, and reset or freeze the EWMA state across halts.

3. Thresholds set from the full sample are look-ahead. “Dollar bars at 1/50 of average daily notional” is a full-sample statistic unless the average is trailing. In a trending-volume market such as crypto 2020–2021, a full-sample threshold produces too few bars early and too many late. The returns look stationary because the clock was tuned with knowledge of the future.

4. The bar timestamp is not the decision time. A bar closes on the trade that crosses the threshold, and you learn about that trade after feed latency. Features stamped at \(T_k\) are fine. Entries at \(p_{\text{close},k}\) are not: they are same-bar execution in a new costume. Using the bar’s VWAP as the entry price is worse, because it averages trades that happened before the signal existed.

5. Asynchronous closes poison cross-asset joins. Volume bars on two assets never close together. Joining their features on bar index \(k\) is look-ahead whenever asset B’s \(k\)-th bar closed after asset A’s. The correct join is an as-of join on timestamps (each row sees only B’s last completed bar). Even done correctly, non-synchronous sampling biases correlations toward zero at short horizons (the Epps effect) and can manufacture lead–lag relationships out of sampling alone. Interpolation-free estimators for asynchronous tick data exist and are worth using before believing a cross-asset signal built on activity bars.

6. Exchange time versus receive time. Activity clocks amplify timestamp errors. A burst of prints with identical exchange timestamps but different receive times can be ordered differently by two pipelines, which changes which trade closes the bar and what the bar’s features are. Pick one timestamp, sort stably, and store both.

The monetization gap: fee-wall arithmetic

Information-driven bars reliably improve three in-sample diagnostics: returns closer to Gaussian, lower heteroskedasticity, and, for imbalance and run bars, higher serial correlation of bar returns. The first two are the mechanical normalization above, and they are real. The third is partly constructed. A bar that closes because one-sided aggression accumulated tends to be followed by continuation (meta-orders keep working) and also by bounce at the close print. Neither is guaranteed to survive being measured at your fill price instead of the last trade.

Now the fee wall. Suppose a feature built on dollar-imbalance bars predicts the sign of the next bar’s return with directional accuracy \(p\), and the next bar’s mean absolute return is \(\mathbb{E}\lvert r_\tau\rvert\). The gross edge per round trip from trading the sign is \((2p-1)\,\mathbb{E}\lvert r_\tau\rvert\), so break-even against a round-trip cost \(C_{\text{rt}}\) requires

Use crypto-perp numbers that are typical for a non-VIP taker in late 2026: 2 × 2.5 bps fees, 1 bp of quoted spread, and 1.5 bps of extra slippage from arriving just after a sweep. That gives \(C_{\text{rt}} = 7.5\) bps. With short bars where \(\mathbb{E}\lvert r_\tau\rvert = 8\) bps, the break-even accuracy is \(p^{*} = 0.5 + 7.5/16 \approx 0.97\). With bars four times longer (\(\mathbb{E}\lvert r_\tau\rvert = 20\) bps) it is still \(0.69\). The published “58% accuracy on imbalance bars” result fails this check by a factor of several.

Stated as \(R^2\): a linear forecast with correlation \(\rho\) to the next return, traded on its sign, earns about \(\rho\,\sigma_\tau\sqrt{2/\pi}\) per trade gross. An excellent-sounding out-of-sample \(R^2 = 2\%\) (\(\rho \approx 0.14\)) on bars with \(\sigma_\tau = 10\) bps earns about \(0.14 \times 10 \times 0.80 \approx 1.1\) bps per trade, roughly a seventh of the 7.5 bps cost. Statistical and economic significance part ways at exactly this point. Run your own horizon and cost assumptions through the transaction-cost break-even calculator before reading anything else in a paper’s results table.

The conditional cost is worse than the unconditional one, and that is the part specific to activity bars. Imbalance-bar closes cluster where liquidity is thinnest. By construction they fire when a one-sided sweep has just completed: the swept side of the book is depleted, quoters have widened or pulled, and every other participant watching the same flow fires at the same moment. Your average spread over the day might be 1 bp. Your spread conditional on a DIB close is the post-sweep spread, often a multiple of that. Market-maker adverse-selection mechanics predict it: the moment your signal fires is the moment the passive side is pricing in the most toxicity. Execution as a maker does not escape it either. Resting at the bar close, you fill only when the continuation fails, which is adverse selection in its purest form. A backtest that charges average costs to trades triggered at bar closes understates costs most where its signal is strongest.

Two smaller accounting traps compound this. Bar counts per day vary, so annualizing a Sharpe ratio as \(\sqrt{\text{bars per year}}\) times the per-bar Sharpe is wrong whenever activity trends. Annualize in calendar time from daily P&L. And a falsification study that applies honest constraints to OHLCV-based intraday signals finds the edge that survives costs is far smaller than the in-sample statistics suggest. That is the expected outcome here, not a surprise.

Which clock for which job

• Volatility and covariance estimation: tick or volume time, with noise-robust estimators. This is where activity clocks earn their keep.
• Daily or slower cross-sectional features: time bars are fine. The variance-mixture problem averages out at daily aggregation, and the joins are simpler.
• Execution scheduling: volume time, because that is what participation-rate and VWAP algorithms actually track.
• Short-horizon event detection: imbalance or run bars as triggers, labelled with costed forward returns measured from realistic fill times.
• Intraday ML datasets: dollar bars with trailing thresholds, as-of joins, and calendar-time evaluation. Benchmark against time bars over the same horizon. If the activity clock’s advantage disappears once costs are charged, the advantage was the mechanical channel.

Production failure modes

• Live bars disagree with backtest bars. Live feeds drop, duplicate and reorder prints, so the live bar sequence diverges from the cleaned historical one within hours. Any feature that depends on which trade closed the bar is not reproducible live. Replay the recorded live feed through the research code and diff the bars weekly.
• Threshold drift without alerts. A trailing-notional threshold quietly halves after a volume regime change, the bar count doubles, and every per-bar feature changes scale. Monitor bars per day as a first-class metric.
• Silent auction contamination. One unfiltered closing-auction print creates a bar whose “return” is the auction imbalance. Filter by sale condition codes.
• Venue fragmentation. Single-venue dollar bars on an asset that trades on twelve venues sample one venue’s share of flow, and that share moves with fee schedules.
• The signal fires with everyone else’s. Imbalance-bar logic is public and widely implemented. Live fills at the trigger show the crowding directly, as markouts worse than the backtest’s assumed cost.

Questions to ask a paper (or backtest)

1. Which clock, and were its thresholds set with trailing or full-sample statistics?
2. How are trade signs obtained (aggressor flag, tick rule, Lee–Ready), and were fragmented prints aggregated?
3. Is the entry price the bar close, the next trade, or a simulated fill after latency?
4. What are the round-trip cost and \(\mathbb{E}\lvert r_\tau\rvert\) at the prediction horizon, and does the claimed accuracy clear \(p^{*}\)?
5. Is the comparison against time bars run at the same calendar horizon, with the same costs?
6. For multi-asset features: as-of joins on timestamps, or joins on bar index?
7. Are the reported normality or autocorrelation improvements used as evidence of tradeable information? They are evidence of a better clock and nothing more.

What our scoring captures

Papers that resample to activity bars score on our empirical-rigor axis mostly according to what they do after resampling: costed evaluation from realistic fill times, trailing calibration, and an honest time-bar baseline. “Dollar bars improve normality” is a property of the clock and earns little. “A DIB-triggered signal survives 2× costs out of sample” is a result, and it is rare. The broader literature on order-flow clocks sits in the market microstructure and econometrics & time series hubs.

Run your numbers: transaction-cost break-even calculator → · tick-data storage sizer → · The data underneath: market data types → · storing tick data → · What the bars are detecting: order-flow intensity and Hawkes processes → · LOB imbalance → · The field: LOB research map →

Background: the classic references are Clark (1973), Ané and Geman (2000), Easley, López de Prado and O’Hara (2012) on the volume clock, and López de Prado (2018, ch. 2) for the imbalance and run bar definitions. Practitioner treatments that informed this synthesis include Trading the Breaking and HangukQuant Research. Browse the flowchart library, the rest of the guides, and the tools.

Definition. The conditional intensity of an event stream is the instantaneous arrival rate given everything observed so far,

where \(N(t)\) counts events (trades, buy market orders, cancellations) up to \(t\). A Poisson process has constant intensity. Order flow does not: one trade raises the probability of the next, through order splitting, herding, liquidity reactions and matching-engine mechanics. Every short-horizon feature built on trades is an estimator of this intensity, whether its author calls it that or not.

The ghost drop-off: why fixed windows lie

The most common order-flow feature is a trailing-window statistic: trade count \(N(t) - N(t-w)\), or signed volume \(\sum_{t-w < t_i \le t} \epsilon_i v_i\) with \(\epsilon_i = \pm 1\) the aggressor side. It weights every event inside the window equally and every event outside it at zero, which is a boxcar kernel. The defect is easy to state. A 500-lot buy at time \(s\) contributes fully to the feature until \(s + w\) and then disappears. The feature jumps at \(s + w\), when nothing happened in the market. Every large print creates a second, phantom event exactly one window-length later. A model trained on the feature learns to react to those ghosts, and the ghosts are perfectly predictable, which also makes them look like signal in a backtest.

Count-based windows (the last \(n\) trades) avoid the calendar ghost but replace it with a variable clock. In a busy market the last 20 trades span 300 milliseconds. In a quiet one they span four minutes, so the “same” feature measures different horizons in different regimes. That is the sampling-clock problem again, at the feature level.

The fix is to make the weight a smooth, monotone function of age, so that no event ever drops off all at once. The exponential kernel is the only choice that also stays memoryless, which is what makes it cheap to compute.

Exponentially decayed trade imbalance

Define the decayed imbalance with a concave volume weight:

The recursion. Because \(e^{-\gamma(t - t_i)}\) factorizes, the sum never needs to be recomputed. On each event \(n\):

This is one multiply and one add per event, with \(O(1)\) state. The second clause is where most implementations go wrong: the decay has to be applied at query time as well. A feature read as “the last updated \(\Lambda\)” freezes during quiet periods. That is a ghost in the opposite direction, where an old burst looks current because nothing has arrived to decay it.

Normalization. A bounded version divides by decayed gross volume. Both sums follow the same recursion:

The pseudo-count \(c\) matters. Without it, a single small trade after a quiet minute sets the ratio to \(\pm 1\), the strongest possible reading, on the weakest evidence the market can produce.

Why \(\alpha = 0.5\). The empirical square-root law says the price impact of a metaorder of size \(Q\) scales as \(I(Q) \approx Y\sigma\sqrt{Q/V}\), with \(V\) daily volume and \(Y\) of order one. If the feature is meant to proxy accumulated price pressure, a trade should contribute in proportion to its expected impact, not its size. \(\alpha = 1\) lets one block print dominate the feature for several half-lives. \(\alpha = 0\) (pure counts) ignores size and is dominated by fragmentation. \(\alpha = 0.5\) borrows the curvature of the impact law and is a heuristic, not a derivation. The square-root law describes metaorders, the sum of many child orders, and per-print impact is usually even more concave. The square-root / order-imbalance reconciliation literature is the place to see how these scales fit together.

There is a consequence that is easy to miss. With \(\alpha < 1\), print aggregation is no longer optional. The square root is subadditive. A single 100-lot sweep reported as 25 prints of 4 lots contributes \(25 \times \sqrt 4 = 50\) units instead of \(\sqrt{100} = 10\), a fivefold overstatement that grows with how finely the venue fragments fills. Aggregate prints by aggressor order (or by identical exchange timestamp and side) before applying \(v^\alpha\).

Choosing \(\gamma\). Order-sign autocorrelation decays roughly as a power law, \(C(\tau) \propto \tau^{-\beta}\) with \(\beta < 1\). This is the long memory usually attributed to metaorder splitting, and its identification from public data is an active question. One exponential cannot fit a power law, but a small bank can: half-lives of 0.5 s, 5 s, 60 s and 10 min, fed jointly to the model, approximate the power law by a sum of exponentials at four state variables’ cost. Choosing a single “best” half-life by backtest is a parameter search like any other.

Hawkes processes: self- and cross-excitation

The Hawkes process makes the clustering explicit. For a single stream with an exponential kernel,

Here \(\mu\) is the exogenous baseline and \(n\) the branching ratio: the expected number of events directly triggered by one event. A cluster seeded by one exogenous event has expected size \(1/(1-n)\), and a fraction \(n\) of all events is endogenous. Estimates of \(n\) close to 1 across many markets started the “markets are near-critical” literature. Treat them with care. A constant \(\mu\) fitted across an intraday U-shaped activity profile pushes \(n\) toward 1, because the model can only explain the open’s activity burst through excitation. Coarse timestamps and latent common drivers do the same, as the analysis of spurious cross-excitation from common signals shows.

The bivariate model. Split events into buy and sell market orders, \(m, n \in \{\mathrm{b}, \mathrm{s}\}\):

\(\boldsymbol\Gamma\) is the branching matrix, and \(\Gamma_{mn}\) is the expected number of type-\(m\) events triggered by one type-\(n\) event. The process is stationary if and only if the spectral radius \(\rho(\boldsymbol\Gamma) < 1\), in which case the mean intensities are \(\bar{\boldsymbol\lambda} = (\mathbf I - \boldsymbol\Gamma)^{-1}\boldsymbol\mu\).

What the entries separate. \(\Gamma_{\mathrm{bb}}\) (buys triggering buys) is where meta-order splitting lives: child orders of one parent arriving in sequence, plus herding and momentum ignition. \(\Gamma_{\mathrm{sb}}\) (buys triggering sells) is the reactive opposite side: market makers and arbitrageurs leaning against a move, stale-quote arbitrage across venues, stop and liquidation reversals. With limit-order events in the model as well (a four-dimensional process over buy MO, sell MO, bid LO and ask LO), liquidity replenishment appears directly as the excitation of ask-side limit orders by buy market orders. The two mechanisms also work on different clocks. Splitting is slow and persistent (seconds to hours, with power-law memory), while reaction is fast (milliseconds to seconds). A single shared \(\beta\) forces them onto one timescale and blurs them together, so use separate \(\beta_{mn}\), sums of exponentials, or the non-parametric calibration approach.

The decomposition that matters for impact. Take the symmetric case \(\boldsymbol\Gamma = \begin{bmatrix} a & c \\ c & a \end{bmatrix}\). Its eigenvectors are total activity \((1, 1)\) with eigenvalue \(a + c\), and imbalance \((1, -1)\) with eigenvalue \(a - c\). Stationarity is governed by \(a + c\). Persistence of signed flow, which is what moves prices, is governed by \(a - c\): one extra buy triggers an expected cumulative net imbalance of \(1/(1 - (a - c))\). A market can be near-critical in activity (\(a + c \approx 1\)) while its imbalance mean-reverts quickly (\(c \approx a\)). The headline branching ratio does not tell you which. Report both modes. Models that separate core orders from reaction flow push this decomposition into a joint account of order flow, impact and volatility.

Estimation, briefly. The exponential-kernel log-likelihood \(\ell = \sum_i \log\lambda(t_i) - \int_0^T \lambda(s)\,ds\) is also \(O(N)\) via the recursion \(R_i = e^{-\beta(t_i - t_{i-1})}(1 + R_{i-1})\). Goodness of fit has a clean test: under a correct model, the compensator increments \(\int_{t_{i-1}}^{t_i}\lambda(s)\,ds\) are i.i.d. Exp(1) (the time-rescaling theorem). A QQ plot of those residuals is the minimum diagnostic for any Hawkes paper. Extensions worth knowing are state-dependent kernels (spread-dependent excitation) and exact phase-type representations that make measurement and prediction the same computation.

Mechanical and informational excitation

As with book imbalance, measured excitation mixes plumbing with information, and the two have different economic meaning.

Mechanical. One sweep printed as 30 trades with microsecond spacing shows up as enormous self-excitation at near-zero lag. Exchange batching and millisecond timestamp truncation create artificial ties. Clock skew between feed handlers creates artificial lead–lag across venues, and queue depletion at one price produces bursts with no new information behind them. All of this inflates \(\Gamma_{\mathrm{bb}}\) and \(\Gamma_{\mathrm{ss}}\) at the shortest lags and dominates naive fits.

Informational. Persistent same-side flow at seconds-to-minutes lags (splitting by someone who knows something, or simply needs to trade a lot) is the component that predicts drift, and it is the one you pay for as a liquidity provider. Separating the two in practice: aggregate prints first, discard or model sub-millisecond structure separately, then read \(\Gamma_{\mathrm{bb}}\) and its kernel shape at economically meaningful lags.

Avellaneda–Stoikov’s \(\kappa\) and where it breaks

The Avellaneda–Stoikov framework assumes fills on a quote at distance \(\delta\) from the reference price arrive as a Poisson process with intensity

and gives a total optimal spread of \(\gamma\sigma^2 (T - t) + \tfrac{2}{\gamma}\ln\!\left(1 + \tfrac{\gamma}{\kappa}\right)\) around an inventory-skewed reservation price. \(\kappa\) is usually fitted by regressing the log of trade volume reaching each depth on that depth. A large \(\kappa\) means flow concentrates near the touch, and a small \(\kappa\) means market orders routinely walk the book. Two practical problems are well known. The fit needs far more data than one quoting interval provides, and noise in \(\hat\kappa\) feeds straight into the spread. Learning \(\kappa\) online has a regret cost even in the idealized model.

The deeper problem is structural. A–S models fill probability and leaves out post-fill drift. In practice \(A\) and \(\kappa\) are state variables: \(A(t)\) scales with the market-order intensity \(\lambda_{\mathrm{b}}(t) + \lambda_{\mathrm{s}}(t)\), and fills at large \(\delta\) occur preferentially during sweeps, which are exactly the toxic events. A \(\kappa\) pooled across calm and burst states treats deep fills as routine revenue when they actually arrive together with the adverse move. The repair is to price the drift in directly. Quote at depth \(\delta\) only when

which means estimating the conditional markout below rather than the unconditional fill curve. Papers that embed quoting in a Hawkes-driven book, or that predict toxicity directly, are attempts at this repair.

Markouts: what the intensity costs you

For a fill at time \(t_f\), price \(p_f\), with \(s_f = +1\) if you bought and \(-1\) if you sold, and mid-price \(m(\cdot)\):

Plot the median and trimmed mean of \(M(\tau)\) on a log-spaced grid (100 ms, 300 ms, 1 s, 3 s, 10 s, 30 s, 60 s), with day-block bootstrap bands, because markouts are heavy-tailed. A maker’s curve starts near the captured half-spread at \(\tau = 0\) and decays. The level it settles at is the net edge before fees. A curve that crosses zero inside your reaction time means the flow is toxic at speeds you cannot respond to.

The diagnostic that ties this guide together is markouts bucketed by intensity state at the fill: quintiles of \(\lambda(t_f)/\bar\lambda\), or of \(\lvert\widetilde{\mathrm{TI}}_\gamma(t_f)\rvert\). Add buckets for fill depth, aggressor size and sweep versus single-level aggression. Toxicity is not uniform, and an unconditional markout curve averages the state you want to quote in with the state you should be out of. The same machinery measures taker entries: a signal’s markout curve, read from your realistic arrival time, shows how much of the predicted move was left when you got there.

The fee wall, in basis points

Maker side. Take an illustrative mid-cap perpetual: quoted spread 4 bps (half-spread captured 2 bps), maker fee 0.5 bps. Suppose 80% of fills arrive in the calm state with a 5-second markout drift of −0.8 bps, and 20% arrive in the top intensity quintile at −6 bps. The pooled drift is \(0.8(-0.8) + 0.2(-6) = -1.84\) bps, so the net per fill is \(2 - 1.84 - 0.5 = -0.34\) bps. The strategy loses money. Pull quotes in the burst state and the calm fills net \(2 - 0.8 - 0.5 = +0.7\) bps. The intensity model has turned a losing book into a profitable one without improving any price forecast. The catch is detection. The burst is announced by the fills that hit you, so the first fill of every cluster is unavoidable. With branching ratio \(n\), clusters average \(1/(1-n)\) events. At \(n = 0.8\) that is 5, so reacting after the first event avoids at most 80% of the toxic fills, and only if your cancel reaches the matching engine before the second one. Latency puts a ceiling on how much of the model’s value you can capture.

Taker side. Suppose decayed imbalance predicts a 1-second mid move of 1.5 bps in the signalled direction. A non-VIP taker pays about \(2 \times 2.5\) bps in fees plus 1 bp of spread round trip, around 6 bps, before latency. Latency costs more on top: if the predicted move accrues as \(g(\tau) = G\,(1 - e^{-\tau/h})\) and your decision-to-fill latency is \(\ell\), you capture only \(e^{-\ell/h} - e^{-\tau/h}\) of \(G\). With \(h = 200\) ms and \(\ell = 50\) ms, 22% of the move is gone before you arrive. Standalone, the signal is several times short of break-even. Its value is in conditioning things you were doing anyway: quote skew, child-order timing, pausing passive orders. The transaction-cost break-even calculator shows the horizon-versus-cost trade-off with your own numbers, and the transaction-cost guide gives priors by asset class.

Production failure modes

• Stale state across feed gaps. After a reconnect, \(\Lambda\) and every Hawkes intensity are wrong, and a gap looks exactly like a quiet market. Decay across the gap, flag it, and suppress quoting until the state has warmed back up.
• Spurious criticality. A constant baseline fitted over a session with strong seasonality reports \(n \to 1\). Use time-varying \(\mu(t)\) or fit within activity-homogeneous windows.
• Unaggregated prints. Self-excitation at microsecond lags and \(\sqrt{v}\) inflation both come from counting matching-engine fills as decisions.
• Your own orders in the stream. Your fills and cancels excite the intensity you are reading. At any meaningful size the model starts measuring you. Tag and exclude your own events, or model them explicitly.
• Live re-fitting of \(\kappa\). Refit on short windows and spreads oscillate with estimation noise. Fit offline, condition on state live.
• Numerical shortcuts. The unnormalized trick of accumulating \(e^{+\gamma t_i}\) overflows within hours at sub-second half-lives. Always use the decaying recursion.
• Cross-venue lead–lag. Intensity on a lagging venue is a delayed copy of the leading venue’s. Non-synchronous lead–lag estimation is a prerequisite before trusting cross-venue excitation.

Questions to ask a paper (or backtest)

1. What is an event: a print, an aggressor order, or a book update? What is the timestamp resolution, and how were ties handled?
2. Kernel family and baseline: exponential or power-law, constant or seasonal \(\mu\)? Is the branching ratio robust to both choices?
3. Is there an out-of-sample likelihood or time-rescaling residual test, or only in-sample fit?
4. Are markouts reported at several horizons, and conditional on intensity state?
5. For market-making results: does the fill model include post-fill drift, and is queue position simulated?
6. What are the round-trip cost and latency at the prediction horizon, and does the edge clear both?

What our scoring captures

Hawkes papers range from pure theory, which lands in the Lab Rats quadrant by design, to calibrated empirical work with goodness-of-fit diagnostics. Our rigor axis rewards the second kind, with extra weight on out-of-sample likelihood and on markout or cost-aware evaluation. The modelling machinery sits in the stochastic calculus and econophysics & complexity method hubs. Applied work concentrates in market microstructure and HFT & execution.

Run your numbers: transaction-cost break-even calculator → · capacity estimator → · The forces underneath: market-making mechanics → · LOB imbalance → · impact and costs → · The clock question: information-driven bars → · The field: market-making research map →

Background: Hawkes (1971); Bacry, Mastromatteo and Muzy (2015) for the finance survey; Avellaneda and Stoikov (2008); Lillo, Mike and Farmer (2005) on order splitting. Practitioner treatments that informed this synthesis include VertoxQuant on point processes and toxicity, HangukQuant Research on decayed imbalance and intensity fitting, and The Quant Stack on execution. Browse the flowchart library, the rest of the guides, and the tools.

Headline numbers

Where the papers went

Papers can sit in up to three hubs, so shares sum to more than 100%. Measured against the trailing year:

• Gaining share: Market Microstructure 16% vs 11% (+5.3 pp); Reinforcement Learning 10% vs 6% (+4.1 pp); HFT & Execution 9% vs 6% (+2.9 pp)
• Losing share: Risk Management 12% vs 17% (−5.4 pp); Machine Learning 15% vs 19% (−4.1 pp); Insurance & Actuarial 5% vs 7% (−2.5 pp)

Theory or evidence?

Holy Grail = high math and high rigor; Lab Rats = theory without data; Street Traders = evidence with light theory; Philosophers = neither (definitions). The three most rigorous papers of the month: Latent Continuum of Regimes in Limit Order Book Dynamics (rigor 9.0); Compiling and forecasting implied volatility surfaces via survey sampling, tempered inverse-probability sequence learning, and causal activity clocks (rigor 9.0); Arithmetic Models for Commodity Futures, Vanilla, and Calendar Spread Options (rigor 9.0).

Top-rated papers of the month

Ranked by rigor-weighted score (60% empirical rigor, 40% math complexity).

1. Latent Continuum of Regimes in Limit Order Book Dynamics — Holy Grail · Math 8.5 · Rigor 9.0 · Market Microstructure, HFT & Execution · code ✓
2. Compiling and forecasting implied volatility surfaces via survey sampling, tempered inverse-probability sequence learning, and causal activity clocks — Holy Grail · Math 8.5 · Rigor 9.0 · Machine Learning, Volatility
3. Arithmetic Models for Commodity Futures, Vanilla, and Calendar Spread Options — Holy Grail · Math 8.0 · Rigor 9.0 · Volatility, Commodities & Energy
4. Exact Phase-Type Point Processes for Certified Measurement and Prediction of Event Streams — Holy Grail · Math 9.0 · Rigor 8.0 · Market Microstructure · code ✓
5. Packets, Transactions and Queues: Design Principles for HFT Systems from a Measurement Study of CME Market Data — Holy Grail · Math 7.5 · Rigor 9.0 · HFT & Execution · code ✓
6. A game-theoretic analysis of the market microstructure in decentralized exchanges with concentrated liquidity — Holy Grail · Math 9.0 · Rigor 8.0 · Crypto & DeFi, Market Microstructure
7. GROMM: Generative Risk-neutral Option Market Models — Holy Grail · Math 9.0 · Rigor 8.0 · Options & Derivatives, Volatility
8. Data-Driven Dynamic Portfolio Selection with Vine Copula Simulation and Reinforcement Learning — Holy Grail · Math 8.5 · Rigor 8.0 · Risk Management, Portfolio Optimization

Most active authors

Authors with two or more papers dated this month (author pages exist for researchers with three or more papers in the archive):

• Doohwi Cha — 4 papers
• Ke-Wei Huang — 3 papers
• Masaaki Fukasawa — 3 papers
• Junkee Jeon — 3 papers
• Stefan Zohren — 3 papers
• Mihai Cucuringu — 3 papers
• Donghwa Seo — 3 papers
• Seunghan Son — 3 papers

Method

Counts cover every paper in the archive whose own date falls in October 2026; late arXiv listings and re-scores can shift a month’s numbers by a few papers after publication. Topic hubs come from a keyword classifier, scores from an LLM-assisted rubric — heuristics, not peer review. The underlying rows are in the downloadable dataset; previous reports are in the reports archive.

The problem: noise or decay?

A deployed strategy enters a drawdown. Two hypotheses are on the table. Under \(H_0\) the return-generating process is unchanged, and this is an expected excursion into the left tail. Under \(H_1\) the process has shifted to a regime with zero or negative mean, through crowding, a fee-schedule change, a venue upgrade that closed a latency gap, or a regime the edge never covered. The costs are asymmetric. A false positive crystallizes the drawdown and misses the recovery. A false negative keeps paying spread and impact into a negative-expectancy process.

The information limit. Distinguishing a daily Sharpe of \(SR_1\) from \(SR_0\) at roughly \(z\) standard errors takes about

For a strategy whose annual Sharpe falls from 1.5 to 0, the daily difference is \(1.5/\sqrt{252} \approx 0.094\), and \(z = 2\) needs about 450 trading days. No algorithm gets around this for a pure change in the mean. Detectors become fast only when the break also changes something easier to see: volatility, tail shape, autocorrelation, or execution telemetry such as markouts and fill rates. The flip side matters as much. A healthy strategy with a daily Sharpe of 0.10 finishes 400 days under water about 2% of the time (\(400 \times 0.10 = 40\) against a standard deviation of 20). One of our simulated healthy paths did exactly that. Before trusting any rule, look at what healthy drawdowns look like with the equity-curve simulator, and check how much track record the claimed Sharpe needs with the minimum-track-record output of the deflated Sharpe calculator.

Why raw equity curves fail changepoint detectors

Cumulative P&L is an integrated process: \(\mathbb{E}[\mathrm{PnL}_t] = \mathrm{PnL}_0 + \mu t\) and \(\operatorname{Var}(\mathrm{PnL}_t) = \sigma^2 t\). Changepoint models assume piecewise-stationary segments. A random walk’s level drifts away from any fixed segment mean, so a detector fed the equity curve either flags “changes” constantly or, once its variance prior has absorbed the drift, never flags anything. The object to model is the step:

Two practical refinements matter as much as the differencing. Normalize by risk, not dollars. Divide by the gross exposure or the target volatility in force when the position was taken. Otherwise a capital increase, which raises dollar variance, looks exactly like a volatility regime change, and the detector fires on your own allocation decision. Feed net returns, not gross. The most common decay mode is costs rising underneath an unchanged signal (see the fee-wall section below). A detector fed gross returns is blind to it by construction.

BOCPD, written out

Adams and MacKay’s Bayesian online changepoint detection (2007) tracks the run length \(r_t\), the time since the last changepoint, through its posterior \(P(r_t \mid x_{1:t})\). With a predictive density \(\pi_t^{(r)} = P(x_t \mid r_{t-1}, x^{(r)})\) for each live run, the joint updates in two branches:

followed by normalization over \(r_t\). With a constant hazard \(H(\tau) = 1/\lambda_{\text{gap}}\), the prior on regime length is geometric with mean \(\lambda_{\text{gap}}\). For a daily strategy, \(\lambda_{\text{gap}}\) between 250 and 1,000 days encodes “regimes last one to four years.”

The observation model. Returns within a regime are modelled as Gaussian with unknown mean and variance under a normal-inverse-gamma prior \((\mu_0, \kappa_0, \alpha_0, \beta_0)\). The sufficient statistics update in closed form for every run:

and the posterior predictive is Student-\(t\):

Work in log space, and prune run lengths whose posterior mass falls below about \(10^{-8}\), or keep only the top \(K\). That makes each update \(O(K)\) instead of \(O(t)\), with no visible change in output.

Caveat on the “Student-t” label. The \(t\) predictive comes from parameter uncertainty, not from fat tails in the data. As a run grows, \(\alpha_n\) grows, the predictive converges to a Gaussian, and a long healthy run becomes more fragile to a single tail event. In our simulation, one −6σ day inside a healthy regime collapsed \(\mathbb{E}[r_t]\) from about 210 to about 1. The detector called a regime change on a single outlier. Two fixes work. Robustify the inputs by winsorizing at a multiple of a trailing robust scale such as the median absolute deviation (a full-sample scale would be look-ahead). With winsorizing at 4× trailing MAD, the same day barely moved \(\mathbb{E}[r_t]\). Or use a genuinely heavy-tailed or divergence-robust likelihood (the β-divergence variants of BOCPD).

Priors. Set \(\mu_0\) from the deflated backtest mean, not the raw one, because the backtest mean is selected and therefore inflated (why). Set \(\alpha_0 > 2\) so the prior variance is finite, choose \(\beta_0\) to match the backtest variance of \(x_t\), and keep \(\kappa_0\) small so the mean prior counts as one or two observations. The detector then asks “is this still the strategy I validated?”, but only after the validation’s optimism has been taken out.

Expected run length, not the changepoint probability

Many implementations alarm on \(P(r_t = 0 \mid x_{1:t})\). With a constant hazard, that quantity equals \(H\) on every step. Divide the changepoint branch by the sum of both branches and \(\sum P(r_{t-1}, x_{1:t-1})\,\pi_t^{(r)}\) cancels. In our simulation it sat at 0.001 to fifteen decimal places for 600 days, through the break. The evidence for a break shows up one step later, as mass moving onto short run lengths \(r_t = 1, 2, \dots\). Even variants that score the new point under the prior turn \(P(r_t = 0)\) into a single-observation outlier test, which is the wrong question for fat-tailed returns.

The useful summary is the whole posterior. Its first moment,

collapses when mass moves from the long run to short ones, and it collapses for a sustained change rather than a single point. That makes it a natural continuous position scalar:

where \(r_{\min}\) is the regime age you need before trusting a regime with full size (60 days below). Continuous sizing beats a binary switch for the same reason volatility targeting beats on/off risk rules: errors cost a fraction of the position instead of all of it.

The flaw: \(\mathbb{E}[r_t]\) measures age, not quality. After a detected break, the new run starts aging, and \(w_t\) climbs back to 1 after \(r_{\min}\) days whether or not the new regime makes money. The run-length posterior says when the process changed. The posterior of the current run’s mean says whether you still want to own it. The scalar needs both:

where \(P(\mu > 0 \mid x_{1:t})\) is mixed over run lengths using each run’s Student-\(t\) marginal for \(\mu\). A smoother alternative for \(q_t\) is the parameter-uncertainty Kelly shrinkage \(t^2/(1+t^2)\) on the current run’s \(t\)-statistic, the same haircut the Kelly fraction calculator applies to an edge estimated with error. Either way, a fresh regime has to earn its size back.

  flowchart LR
  A["Trade fills"] --> B["Mark-to-market P&L per interval"]
  B --> C["Stationary returns: difference, net of costs, divide by risk"]
  C --> R["Robustify: winsorize at k x trailing MAD"]
  R --> D["Student-t BOCPD posterior P(r_t | x_1:t)"]
  D --> E["Expected run length E[r_t]"]
  D --> Q["Current-regime quality P(mu > 0)"]
  E --> W["Position scalar w_t = min(1, E[r_t]/r_min) x q_t"]
  Q --> W
  C --> K["Conformal lower band on h-step P&L"]
  K --> W
  M["Execution telemetry: markouts, fill rates"] --> W
  W --> S["Sizing with rebalance bands"]

Calibrating it: false alarms versus detection delay

The hazard, robustification and \(r_{\min}\) are the knobs. The only defensible way to set them is to simulate healthy paths and broken paths matched to your strategy’s moments and look at both error types. Statistical process control calls these the in-control and out-of-control average run lengths. We ran a small version: 600 days of Student-\(t\) (\(\nu = 4\)) daily returns at a daily Sharpe of 0.10 (about 1.6 annualized), with a break at day 400 to either a cliff (mean −0.10σ, volatility ×2) or a decay (mean −0.06σ, volatility unchanged). Detector: the NIG BOCPD above, \(r_{\min} = 60\), 60 paths per cell.

Three conclusions hold across the table. The age-only scalar barely protects anything. It avoids 4–11% of post-break exposure, because the losing regime ages back to full size. The quality gate does the protecting. It catches 77–80% of cliff-edge exposure under every setting. The cost is healthy exposure: 12–31% of a good strategy’s sizing is lost to its own noise. Hazard and winsorization move you along a frontier and do not lift it. Sensitive settings catch gradual decay (66%) and pay for it in healthy exposure (69%). Robust settings keep 88% of healthy exposure and catch only 30% of the decay. Run as a binary alarm (kill when \(\mathbb{E}[r_t] < 60\)), the raw detector false-alarmed on 89–98% of healthy 500-day paths. Winsorizing at 4× and 3× trailing MAD with \(H = 1/1000\) brought that down to 33% and 11%, at the price of missing 18–30% of variance-expanding breaks and 65–87% of pure mean decays within 200 days. Which point on the frontier is right depends on how long your strategies usually stay healthy relative to how long a broken one would otherwise keep running. That ratio is a business fact you should estimate, not a statistical constant.

Distribution-free checks: conformal bands, RANSAC trends, robust PELT

BOCPD is model-based. Two complementary tools make fewer assumptions and catch different failures.

Conformal-RANSAC. RANSAC estimates the equity curve’s drift from the largest consensus set. It repeatedly fits a line through a minimal random subset, counts the observations within \(\pm\delta\) of it, keeps the line with the most inliers, and refits on them. Its breakdown point is far above least squares, so a few windfall days cannot manufacture a positive slope, and the inlier ratio is itself a health statistic: the share of history consistent with the dominant drift. Pair it with split-conformal prediction to get a finite-sample lower band. Calibrate the residual quantile \(\hat q\) at level \(\lceil (n+1)(1-\alpha) \rceil / n\) on held-out data, and flag when realized P&L falls below the band. The guarantee \(P(Y < \hat y - \hat q) \le \alpha\) needs only exchangeability. That is the catch. Residuals of a line fitted to an equity curve’s level are not exchangeable. They are the residuals of a random walk, so they are strongly autocorrelated, and the nominal coverage does not hold. Apply the conformal step to \(h\)-step P&L increments instead: predict \(\mathrm{PnL}_{t+h} - \mathrm{PnL}_t \approx \hat\beta h\) using the robust slope, calibrate on past realized \(h\)-step increments (which are close to exchangeable within a regime), and use adaptive conformal inference to adjust \(\alpha_t\) online when coverage drifts. Recent work applies the same idea to regime-weighted conformal VaR control. A practical rule stacks two conditions: robust slope \(> 0\) and current P&L above the conformal lower band. Expect them to disagree during recoveries, when the slope is positive while P&L is still below the band, and decide in advance which one wins.

Robust PELT. Pruned Exact Linear Time segmentation minimizes \(\sum_{\text{segments}} \mathcal{C}(x_{\text{seg}}) + \beta_{\text{pen}} \cdot \#\text{changepoints}\) exactly, with a pruning rule that keeps it near-linear. Replacing squared error with Tukey’s biweight loss (fitted by iteratively reweighted least squares) stops outliers from creating spurious segments. PELT is an offline tool, and it should stay one. Re-running it on an expanding window can move past changepoints as new data arrive, so using it as a live switch reintroduces hindsight. Its real jobs are post-mortems (“when did this actually break?”) and calibrating the BOCPD hazard from the empirical segment lengths across your book. Pre-whiten or aggregate to weekly data first: trend strategies’ autocorrelated P&L over-segments, and mean-reversion strategies’ negative autocorrelation does the opposite. Related corpus work includes Bayesian decision rules that trade false alarms against delay, statistical jump models for regime-based de-risking, Bayesian change-point detection on order flow, and measures of strategy-decay risk.

The fee wall: what a kill switch costs and what it misses

Decay usually arrives as cost creep. Take a short-horizon strategy with a gross edge of 6 bps per trade and 4 bps of round-trip costs: 2 bps net. A competitor arrives and your conditional markouts worsen by 1.5 bps. Net edge falls to 0.5 bps, a 75% decline, while the signal’s directional accuracy and gross information coefficient stay exactly where they were. A monitor built on hit rate, IC or gross \(R^2\) never fires, because the signal never stopped predicting. The prediction just no longer pays for the trip across the spread. This is the same fee wall that order-flow signals and LOB imbalance run into, moving underneath a live strategy. Feed the detector net returns, and monitor the cost components (markouts, not P&L) as leading indicators. Markouts are measured per fill, so they reach statistical significance orders of magnitude faster than daily P&L does.

Whipsaw is a turnover cost. Every off-and-on cycle liquidates and rebuilds the book. At 5 bps per side on 100% gross, a cycle costs about 10 bps of capital before counting any missed recovery. Eight false cycles a year is 80 bps, a fifth of a strategy netting 4% a year. The continuous scalar with rebalance bands (trade only when \(w_t\) moves by more than about 0.2) removes most of this churn. Model the trade-off with the transaction-cost break-even calculator.

Inertia is a drift cost. A dead strategy running at an annual Sharpe of −0.5 and 10% volatility loses about 5% a year. A 150-day detection delay is about 3% of capital. Compare that with the whipsaw cost of the settings that would have caught it sooner. That comparison, not a significance threshold, is the decision.

Production failure modes

• Equity curve in, nonsense out. Feeding levels instead of differences, or dollars instead of risk-normalized returns, makes the detector respond to the integration and to your own allocation changes.
• Gross returns hide cost decay. The commonest death of a live strategy is invisible to a detector that never sees costs.
• Outlier-triggered resets. One fat-tailed day zeroes a naive detector’s run length. Robustify with trailing statistics only.
• Re-armed into a dead regime. Age-only scalars return to full size in a losing regime. Gate on regime quality.
• Hindsight in the switch. Offline segmentation re-run live, full-sample winsorization scales, or priors fitted to the full live history all leak information from the future.
• No state persistence. The run-length posterior is path state. Persist it, and log every update so that a switch-off decision can be reconstructed and audited.
• The human override. A switch that can be overridden “just this once” is a suggestion. Decide the override policy before the first drawdown, in writing.

Questions to ask a paper (or backtest)

1. Is the detector run on differenced, risk-normalized, net returns, or on an equity curve?
2. What are the in-control false-alarm rate and the detection delay, measured on simulated healthy and broken paths matched to the strategy?
3. Is the alarm \(P(r_t = 0)\), which is uninformative under a constant hazard, or a run-length summary?
4. Does the sizing rule gate on the current regime’s quality, or only on regime age?
5. Are robust scales, priors and conformal calibration sets strictly trailing?
6. For conformal claims: on what object is exchangeability assumed, and is it plausible there?
7. Was the switch evaluated on its costed effect (whipsaw turnover and delay drift), or only on classification accuracy?

Related reading in the corpus: hyperbolic alpha decay under crowding, drawdown control with a restart mechanism, and Bayesian ensemble breakpoint detection on energy prices. The method hubs are Bayesian methods and econometrics & time series. The topic hub is risk management. The trade-level counterpart to this essay is exit policy design and ensembling.

Run your numbers: equity-curve simulator → · deflated Sharpe & minimum track record → · Kelly fraction under uncertainty → · volatility-targeting simulator → · The foundations: regime dependence → · confidence intervals for strategy research → · the production checklist →

Sources and background: Adams and MacKay (2007) for BOCPD; Killick, Fearnhead and Eckley (2012) for PELT; Fearnhead and Rigaill (2019) for changepoints with outliers; Gibbs and Candès (2021) for adaptive conformal inference. Practitioner treatments that informed this synthesis are the switch-off series in Trading the Breaking (BOCPD, conformal-RANSAC and robust changepoint protocols). The simulation study is our own. Browse the flowchart library, the guides, and the tools.

1. On the continuity of conditional distributions and stratification for derivative pricing

Holy Grail · Math 8.5/10 · Rigor 7.0/10 · Options & Derivatives, Machine Learning

Stratification can reduce the estimator variance in Monte Carlo simulations by partitioning the sample space according to a low-dimensional factor and sampling separately within the resulting strata, making it well suited to the valuation of path-dependent claims in high-dimensional financial models…

2. Risk Adequacy and Anti-Discrimination in Insurance Pricing: Classification, Aggregation, and Transfers

Holy Grail · Math 8.5/10 · Rigor 6.0/10 · Insurance & Actuarial

Insurance tariffs must reconcile risk-based pricing, portfolio coverage, and anti-discrimination requirements. We introduce a framework in which tariffs are constructed through classification, aggregation of benchmark risk prices within classes, and balanced transfers, and are evaluated by several d…

3. Optimal collective investment under regulatory uncertainty

Holy Grail · Math 8.5/10 · Rigor 6.0/10 · Risk Management, Insurance & Actuarial

We study optimal collective investment subject to a Value-at-Risk (VaR)-type constraint that accounts for uncertainty about future regulatory requirements. The constraint incorporates several possible wealth thresholds and imposes a minimum weighted average probability of meeting them. We characteri…

4. Financial Context-Rich Strategy for Actor-Critic based Agents: A Deep Reinforcement Learning Approach for Stock Trading

Holy Grail · Math 6.5/10 · Rigor 6.0/10 · Reinforcement Learning, NLP & LLMs

In modern financial markets, more than 3/4th of the large volume trade is executed by algorithmic trading based on advanced trading strategies to attain higher precision and speed than traditional human traders. The conventional strategies focused on price-based indicators are inadequate in capturin…

5. Duality of Shortfall Aversion

Lab Rats · Math 9.0/10 · Rigor 2.0/10 · Stochastic Control

Shortfall aversion describes a form of consumption habit formation in which an individual’s utility depends on consumption relative to the historical consumption peak. This study develops a duality framework for a broad class of utility functions, with and without discounting, avoiding the need for…

Browse everything new in the flowchart library, or explore by area on the topic hubs.

Authors: Raphael Coelho

Abstract

When two forecasts of a tail, such as prediction intervals at high coverage, extreme quantiles or expected shortfall (ES), are compared by a consistent score, the score difference is known in advance except on the days on which the outcome falls beyond a forecast. The number of these exceedances, not the number of observations, is the effective sample size. For intervals at miscoverage level $α$, the number of observations a comparison needs grows like $1/α$ while the number of exceedances it needs does not, and for some pairs of forecasts no test can do better. A confidence sequence that is valid at all times and needs no distributional assumption attains this rate; Hoeffding-type procedures, calibrated by the range of the score difference, need $1/α$ times more. When exceedances are scarce, the Diebold-Mariano test declares the less conservative forecaster better with at least, asymptotically, the probability of observing no exceedance, whichever forecaster is actually better; the usual small-sample corrections do not remove this, while a bootstrap-t version largely does. For (value-at-risk, ES) forecasts, a less conservative ES can be shown without assumptions to be worse on the days on which it is less conservative, but never to be better without an assumption on the loss tail. In 250-day windows of 97.5% forecasts for seven stock indices, 18.4% of significant Diebold-Mariano results favouring the less conservative model contradict the ranking on the remaining days, against 6.9% of those favouring the more conservative one.

Complexity vs Empirical Score

• Math Complexity: 8.0/10
• Empirical Rigor: 7.0/10
• Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: This paper presents a highly mathematical and rigorous analysis of tail forecast comparison, introducing novel insights into effective sample size and Diebold-Mariano test limitations. The empirical sections, while present, could benefit from more explicit data and code availability for full reproducibility, but the methodology is well-detailed.

Research Flowchart

  flowchart TD
    A[Research Goal: Compare Tail Forecasts & Assess Sample Size Impact] --> B{Methodology: Consistent Score Comparison & Effective Sample Size}
    B --> C[Data/Inputs: Tail Forecasts (Prediction Intervals, Extreme Quantiles, ES)]
    C --> D{Computational Processes: Confidence Sequences, Hoeffding-type Procedures, Diebold-Mariano Test, Bootstrap-t}
    D --> E[Empirical Analysis: 250-day windows, 97.5% forecasts, 7 Stock Indices]
    E --> F[Key Finding 1: Effective Sample Size = Number of Exceedances (not observations)]
    F --> G[Key Finding 2: Diebold-Mariano test issues with scarce exceedances; Bootstrap-t helps]
    G --> H[Key Finding 3: Less conservative ES can be worse without loss tail assumption]
    H --> I[Key Finding 4: DM results favouring less conservative model often contradict other days' ranking (18.4% vs 6.9%)]

Authors: Fail Gafarov

Abstract

We study coordinated groups of wallets on a bonding-curve launchpad (pump.fun on Solana), using a complete record of 200,443 token launches, 15,854,121 curve trades and 51,025,958 SOL transfers over 88.3 hours. A prospective, multi-channel discovery procedure (co-buying, shared funding, joint signing, transfers, consolidation) yields 23,710 wallet groups. Without ground-truth labels we validate them by out-of-sample link prediction, independent non-trading links and a comparison with naive co-buy clustering at equal numbers of flagged tokens: raw clustering collapses into one giant component, yet it ranks tokens at least as well as the groups, so the value of the structure is descriptive. Group activity is highly concentrated and splits into serial creators and long-lived rings of early buyers. Known members buy earlier than other buyers of the same token and extract more cash than an activity-matched placebo set, while late buyers’ net cash flow is negative; tokens with members show a pump-like price path, but the contrast vanishes when controls are restricted to comparable tokens, so we report it as an association. A follower entering at the earliest group signal loses money out of sample in all 32 configurations and does no better than random entry. A few pools of tens to over a hundred fast wallets are associated with a two- to four-fold higher chance of reaching 20 SOL, also within creators, with a smaller excess at 60 SOL and none for graduation; on a final hold-out the team rules keep a precision of about one half to three quarters, yet many signals arrive when the curve already holds 20 SOL, one team supplies most of them, and they add no information beyond early curve state. We also document evaluation errors that turn the same data into apparent profit.

Complexity vs Empirical Score

• Math Complexity: 5.0/10
• Empirical Rigor: 9.0/10
• Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper demonstrates strong empirical rigor through extensive data analysis and robust validation methods for wallet groups. Its novelty lies in the multi-channel discovery procedure and the detailed, population-level study of coordinated activity on a specific blockchain launchpad. While the mathematical complexity is moderate, the statistical methodologies are advanced and well-applied.

Research Flowchart

  flowchart TD
    A[Research Goal: Study Coordinated Wallet Groups on pump.fun] --> B{Data Collection: 200k launches, 15M trades, 51M SOL transfers (88.3 hrs)}
    B --> C[Methodology: Multi-channel Wallet Group Discovery (co-buying, funding, signing, transfers)]
    C --> D{Computational Process: Yields 23,710 Wallet Groups}
    D --> E[Validation: Out-of-sample link prediction, non-trading links, comparison with naive clustering]
    E --> F{Key Findings: Group Activity & Impact}
    F --> G1[G1: Groups split into serial creators & long-lived early buyer rings]
    F --> G2[G2: Members buy earlier, extract more cash (late buyers lose money)]
    F --> G3[G3: Tokens with members show pump-like path (association, vanishes with controls)]
    F --> G4[G4: Follower entering at earliest group signal loses money]
    F --> G5[G5: Fast wallets associated with higher chance of reaching 20 SOL (precision 1/2-3/4)]

Authors: Danila Litvinov

Abstract

We establish multidimensional exponential convergence of the lifted Bass gradient flow in martingale optimal transport. Let $μ\le_{cx}ν$ be an irreducible pair, with compact convex target support, source support strictly inside the target support, and target density bounded above and bounded away from zero almost everywhere on its interior. For every bounded $σ(X)$-measurable initial state, the flow converges exponentially in $L^2$ to the global minimizer with the prescribed mean, and the functional value gap decays exponentially. Both bounds hold at every nonnegative time with the exact initial errors as prefactors. The quantitative step is strong convexity on sets with a common $L^\infty$ bound and a fixed mean. We prove it by interpolating entropic endpoint potentials, using two conditional copies to cancel the unknown potential increment, and obtaining a Hessian lower bound uniform in the regularization parameter. Removing the regularization yields curvature for the original functional; the known qualitative theory supplies a common bound for the orbit. This extends the one-dimensional exponential-convergence result of Backhoff-Veraguas, Pammer, and Schachermayer. The main result is formalized in Lean 4. AI was used in preparing the paper (see Section 8).

Complexity vs Empirical Score

• Math Complexity: 9.0/10
• Empirical Rigor: 2.0/10
• Quadrant: Lab Rats — theoretically deep, empirically untested

Why this score: The paper presents a highly theoretical and mathematically dense analysis of exponential convergence for a specific gradient flow. While the mathematical rigor and novelty are strong, there is no empirical validation or backtesting, placing it firmly in the ‘Lab Rats’ quadrant. The formal verification in Lean 4 significantly boosts reproducibility.

Research Flowchart

  flowchart TD
    A[Research Goal: Establish Multidimensional Exponential Convergence of Lifted Bass Gradient Flow] --> B{Key Methodology: Strong Convexity & Interpolating Entropic Endpoint Potentials};
    B --> C[Inputs: Irreducible Pair μ ≤_cx ν, Compact Convex Target Support, Source Support < Target Support, Target Density Bounded];
    C --> D[Computational Processes: Lean 4 Formalization, AI for Paper Preparation];
    B --> E{Key Method Detail: Using Two Conditional Copies to Cancel Unknown Potential Increment};
    E --> F[Quantitative Step: Hessian Lower Bound Uniform in Regularization Parameter];
    F --> G[Outcomes: Exponential L2 Convergence to Global Minimizer, Exponential Functional Value Gap Decay, Bounds with Exact Initial Errors];
    A --> G;

Authors: Szymon Lis, Robert Ślepaczuk, Paweł Sakowski

Abstract

This study examines whether transformer-based financial news sentiment improves equity volatility forecasts beyond conventional market-based predictors. Using data for eleven US stocks, we compare volatility persistence models with specifications incorporating the VIX, AAII investor sentiment, and news sentiment extracted using ModernFinBERT. The forecasting target is annualized volatility constructed from future averages of trailing 252-session Yang-Zhang estimates, evaluated at horizons of 1, 5, and 22 trading days. Models are assessed through walk-forward testing with expanding and rolling estimation windows, fixed initial hyperparameter selection, and statistical comparisons accounting for multiple testing and serial dependence. Results show that a naive persistence benchmark achieves the lowest average QLIKE loss at all horizons. Although HAR-VIX produces the largest number of individual forecasting wins under expanding windows, it does not achieve the lowest mean loss. Adding text-based sentiment improves HAR forecasts in only two of 33 expanding-window comparisons, and neither improvement remains statistically significant under longer HAC bandwidths. A blinded audit of 2,700 news texts further characterizes the agreement and confidence patterns of ModernFinBERT classifications. The findings do not establish a general incremental forecasting advantage from transformer-based news sentiment. They highlight the importance of strong benchmarks, temporal alignment, robust inference, and careful evaluation of text-derived predictors.

Complexity vs Empirical Score

• Math Complexity: 6.0/10
• Empirical Rigor: 8.0/10
• Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper employs advanced statistical methods for volatility forecasting and rigorously tests hypotheses across multiple assets and horizons. While the core methodology builds on established models, the application of transformer-based sentiment and the thorough comparative analysis contribute to its novelty. The findings, though negative, are robust and well-supported.

Research Flowchart

  flowchart TD
    A[Research Goal: Transformer-Based Sentiment for Equity Volatility Forecasts?] --> B(Methodology: Compare Volatility Models);
    B --> C{Data/Inputs: 11 US Stocks};
    C --> D[Computational Processes: Walk-Forward Testing];
    D --> E(Model Comparison: QLIKE Loss & Statistical Tests);
    E --> F[Key Findings: No General Incremental Forecasting Advantage];
    F --> G(Outcome: Highlight Importance of Benchmarks & Robust Inference);

    C -- Data Details --> C1(Yang-Zhang Volatility, VIX, AAII, ModernFinBERT Sentiment);
    D -- Process Details --> D1(Expanding & Rolling Windows, Fixed Hyperparameters);
    E -- Comparison Details --> E1(Naïve Benchmark, HAR-VIX, HAR + Sentiment Models);

Authors: Weijia Han, Lisha Qu

Abstract

Factor models are central to asset pricing and portfolio construction, but neural factors are often difficult to interpret, and their effect on portfolio risk is unclear. A sparse autoencoder (SAE) may address both concerns by describing daily returns across many assets with only a few active features. We ask whether this sparse representation produces stable financial factors and better-diversified portfolios. We study cryptocurrency returns, use equities as a secondary test, and compare the SAE with common benchmarks in a time-ordered evaluation. We find consistently lower volatility in SAE portfolios and link this result to lower correlations among the selected assets. However, the SAE does not outperform a simple contrarian strategy on risk-adjusted return, and its factor identities are not stable across training runs. Our contribution is to show that sparse coding can support portfolio diversification, while clearly separating this risk effect from claims about alpha or stable factor identities.

Complexity vs Empirical Score

• Math Complexity: 6.0/10
• Empirical Rigor: 6.0/10 (capped from 8.0: transaction costs not modelled; see penalty flags)
• Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper presents a solid empirical study using sparse autoencoders for portfolio diversification, clearly outlining its methodology and comparing against relevant benchmarks. While the mathematical complexity is moderate, the rigorous empirical evaluation across multiple asset classes and careful consideration of biases contribute significantly to its quality. The novelty lies in applying monosemanticity-style SAEs to financial factor discovery, specifically for risk reduction rather than alpha generation.

Research Flowchart

  flowchart TD
    A[Research Goal: Do SAEs produce stable financial factors & better-diversified portfolios?] --> B{Methodology: Compare SAE with common benchmarks};
    B --> C[Data: Cross-sectional Cryptocurrency Returns (Primary) & Equities (Secondary)];
    C --> D{Computational Process: Time-ordered evaluation of factor models & portfolio construction};
    D --> E[Outcome 1: Consistently lower volatility in SAE portfolios due to lower asset correlations];
    D --> F[Outcome 2: SAE does not outperform simple contrarian strategy on risk-adjusted return];
    D --> G[Outcome 3: SAE factor identities are not stable across training runs];
    E & F & G --> H[Conclusion: SAE supports portfolio diversification, but not alpha or stable factor identities];

Authors: Antoine Jacquier

Abstract

For one-dimensional strict local martingales of the form $\mathrm{d} X_t=σ(X_t) \mathrm{d} W_t$ starting from $X_{0}=x_{0}>0$, we prove the strong rate $\mathbb{E}\left[|X^h_N-X_T|\right] \lesssim h^{\min{θα,\frac{1}{2}}}$ for a discrete scheme~$X^h$ truncated at a level $h^{-α}$, with time step~$h$, as $h\downarrow 0$, where~$θ$ depends on the truncation property. For the CEV model $σ(x)\equiv x^{1+γ}$ in particular, the rate is~$\frac{1}{2}$ for any $γ\in (0,1)$ (up to a logarithmic pre-factor at $γ=\frac{1}{2}$).

Complexity vs Empirical Score

• Math Complexity: 9.0/10
• Empirical Rigor: 4.0/10
• Quadrant: Lab Rats — theoretically deep, empirically untested

Why this score: This paper is highly theoretical, focusing on mathematical proofs for strong convergence rates of a numerical scheme. While it includes a ‘Numerical validation’ section, the empirical rigor is limited to demonstrating the theoretical results rather than extensive backtesting or real-world data application. The novelty lies in establishing the first strong rate for discretisation schemes of strict local martingales.

Research Flowchart

  flowchart TD
    A[Research Goal: Strong Rate for 1D Driftless SLMs] --> B{Key Methodology: Truncated Discrete Scheme};
    B --> C[Model/Data: dX_t = σ(X_t) dW_t, X_0=x_0>0];
    C --> D{Computational Process: Analyze E[|X^h_N - X_T|] as h -> 0};
    D --> E[Outcome 1: Strong rate E[|X^h_N - X_T|] ~ O(h^min{θα, 1/2})];
    E --> F[Outcome 2: For CEV model σ(x)=x^(1+γ), rate is 1/2 for γ in (0,1)];

Authors: Bernhard K Meister

Abstract

The term Kessler syndrome was coined for a runaway orbital cascade creating a high-density debris field. The idea is generalised and applied to supply chains. An additive channel enables failures to generate further failures; a subtractive channel erodes the capacity to contain and restore. When damage outpaces repair, failure becomes irreversible and the system settles into a new state. Supplies connect the theory to its application. Stocks insulate subsystems from one another, and their exhaustion exposes dependencies. For the Persian Gulf states, restrictions on maritime supplies from spring 2026 are inadequately compensated by air and road transport alternatives. This depletes inventories unevenly across products and impairs the functioning of the logistics system itself. Throughput optimisation builds fragility everywhere; the Gulf is where it is becoming visible, with the Strait of Hormuz restriction as the trigger. Formulated as a Kelly portfolio allocation under a weakest-link payoff, the optimal strategy concentrates resources on the lagging rather than the leading inputs.

Complexity vs Empirical Score

• Math Complexity: 6.0/10
• Empirical Rigor: 3.0/10
• Quadrant: Lab Rats — theoretically deep, empirically untested

Why this score: The paper introduces a novel conceptual framework with some mathematical formalization (Kelly criterion, Leontief production) but lacks empirical validation or detailed data analysis. The focus is on theoretical generalization and application to a specific scenario rather than rigorous backtesting.

Research Flowchart

  flowchart TD
    A[Research Goal: Generalise Kessler Syndrome to Supply Chains & Apply to Persian Gulf] --> B{Key Methodology: Generalised Kessler Syndrome Concept}
    B --> C[Data/Inputs: Supply Chain Dynamics, Maritime Restrictions (Spring 2026), Persian Gulf Logistics, Stock Levels]
    C --> D{Computational Process: Kelly Portfolio Allocation, Weakest-Link Payoff}
    D --> E[Key Finding 1: Gulf Supply Chain Fragility due to Throughput Optimisation]
    E --> F[Key Finding 2: Strait of Hormuz Restriction as Trigger for Irreversible Failure]
    F --> G[Key Finding 3: Optimal Strategy: Concentrate Resources on Lagging Inputs (Kelly Portfolio)]

Authors: Yadh Hafsi, Fabrizio Lillo, Grégoire Szymanski

Abstract

We develop a framework for calibrating passive market impact, defined as the price response to liquidity-providing limit orders, in contrast to the impact of liquidity-consuming market orders. We build on the model of Ouazzani Chahdi, Rosenbaum, and Szymanski (arxiv:2412.07461), which combines Hawkes-driven market-order flow with queue-reactive limit-order and cancellation dynamics. Within this model, we show that the price decomposes into a universal propagator term driven by signed order-flow imbalance and a state-dependent correction that captures the dependence of price sensitivity on volume at the best quotes. This decomposition links limit-order-book microstructure to standard propagator models and yields a regression on two observable signals: the propagator signal and the queue-weighted order-flow imbalance. We calibrate the model on high-frequency order-book data and find that the passive-impact term explains substantial variation in prices beyond the classical propagator, improving both model fit and the accuracy of predicted impact curves.

Complexity vs Empirical Score

• Math Complexity: 8.5/10
• Empirical Rigor: 8.0/10
• Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: This paper presents a highly sophisticated mathematical framework for passive market impact, building on advanced Hawkes processes and queue-reactive dynamics. It demonstrates strong empirical rigor through calibration on high-frequency order-book data and improved model fit, offering a novel decomposition of price impact.

Research Flowchart

  flowchart TD
    A[Research Goal: Calibrate Passive Market Impact] --> B{Key Methodology: Extend OCS Model (Hawkes Flow + Queue Dynamics)};
    B --> C[Model Development: Decompose Price into Propagator + Queue-Sensitive Correction];
    C --> D[Data/Inputs: High-Frequency Limit Order Book Data (Equities)];
    D --> E[Computational Process: Regression on Propagator Signal & Queue-Weighted Order-Flow Imbalance];
    E --> F[Key Finding 1: Passive Impact Explains Substantial Price Variation Beyond Classical Propagator];
    E --> G[Key Finding 2: Improved Model Fit and Accuracy of Predicted Impact Curves];

Authors: Hanchao Liu, Dena Firoozi

Abstract

We develop a mean field game model for optimal trading in which the interaction among traders is mediated by the dynamics of the limit order book. The bid- and ask-side order densities are modelled as stochastic processes with values in infinite-dimensional function spaces that solve stochastic partial differential equations (SPDEs) with stochastic order submission rates and multiplicative infinite-dimensional noise. This extends an existing limit order book framework to capture the aggregate impact of the strategic traders on the order densities. Trading activity affects the order densities, which determine the market depths near the best quotes and hence the mid-price, and the mid-price in turn affects the traders’ decisions. We establish the well-posedness and regularity of the resulting density-price system and use it to formulate an $N$-player optimal trading game in which the traders’ market orders enter the book through their aggregate order flow. We then study the corresponding mean field game as the number of traders tends to infinity, characterize the mean field equilibrium as a fixed point of the associated consistency condition, and prove its existence and uniqueness under suitable conditions. Finally, we show that the mean field equilibrium induces an $ε_N$-Nash equilibrium for the finite-player game, with $ε_N=O(N^{-1/2})$, and provide a numerical illustration of the equilibrium dynamics.

Complexity vs Empirical Score

• Math Complexity: 9.0/10
• Empirical Rigor: 3.0/10
• Quadrant: Lab Rats — theoretically deep, empirically untested

Why this score: This paper presents a highly complex mathematical framework using SPDEs and Mean Field Games to model optimal trading within a limit order book. While the theoretical development is rigorous and novel, the empirical validation is limited to a numerical illustration, lacking extensive backtesting or real-world data analysis. The overall contribution is strong due to its innovative theoretical approach.

Research Flowchart

  flowchart TD
    A[Research Goal: Model Optimal Trading with SPDE-based LOB & MFG] --> B{Key Methodology: Extend LOB to SPDE, Formulate N-Player Game, Study MFG};
    B --> C[Inputs: Stochastic Order Submission Rates, Multiplicative Infinite-Dimensional Noise];
    C --> D{Computational Process: Well-posedness & Regularity of Density-Price System, Fixed Point for MFG Equilibrium};
    D --> E[Outcomes: Characterize MFG Equilibrium (Fixed Point), Prove Existence & Uniqueness];
    E --> F[Further Outcomes: MFG Induces ε_N-Nash Equilibrium (ε_N=O(N^-1/2)), Numerical Illustration of Dynamics];

Authors: Kaname Imagawa

Abstract

This paper investigates the mechanisms and conditions underlying negative convexity in pass-through mortgage-backed securities (MBSs) backed by fixed-rate mortgages. We first derive an exact decomposition of MBS prices into three components: the present value of interest-rate-independent cash flows, the present value of principal repaid early through interest-rate-dependent prepayments, and an interest rate option component capturing the reduction in future cash flows due to prepayments. Under a first-order approximation, the latter two components resemble portfolios of floorlets and Asian-style interest rate options, respectively. We then apply a cumulant expansion to derive closed-form valuation approximations under the one-factor Hull–White model. Numerical results show that including the second cumulant improves accuracy relative to a first-cumulant approximation. Building on the price decomposition, we decompose duration and convexity into component-level contributions and express the condition for negative convexity in terms of their second-order interest rate sensitivities. Finally, phase diagrams across interest rate levels and refinancing sensitivities identify the regions where negative convexity occurs and the corresponding sign boundaries. The negative-convexity region expands as the long-run mean of refinancing sensitivity increases. Negative convexity can arise even when no refinancing incentive exists initially, while its magnitude may vary nonmonotonically with the interest rate level. These findings explain negative MBS convexity through the sensitivities of its price components and quantitatively characterize the conditions governing its occurrence.

Complexity vs Empirical Score

• Math Complexity: 8.5/10
• Empirical Rigor: 4.0/10
• Quadrant: Lab Rats — theoretically deep, empirically untested

Why this score: The paper demonstrates high mathematical sophistication through its derivations, cumulant expansion, and decomposition of MBS prices and sensitivities. However, it lacks empirical backtesting or extensive data validation, focusing primarily on theoretical and numerical results. The novelty lies in the exact decomposition and the application of cumulant expansion for closed-form approximations.

Research Flowchart

  flowchart TD
    A[Research Goal: Understand Negative Convexity in Pass-Through MBS] --> B{Key Methodology: Price Decomposition & Valuation Model};
    B --> C[Data/Inputs: Fixed-Rate MBS, Interest Rate Levels, Refinancing Sensitivity];
    C --> D{Computational Processes: Cumulant Expansion, Numerical Valuation, Duration/Convexity Decomposition};
    D --> E[Key Findings: Negative convexity arises from component sensitivities; Regions identified via phase diagrams; Magnitude non-monotonic with interest rates; Expands with long-run mean refinancing sensitivity];

Authors: Dong Yan, Haocheng Zhu, Yao Li

Abstract

This paper proposes a deep learning-driven operator splitting scheme for high-dimensional free-boundary problems, with American option pricing as a representative application. In a market with proportional transaction costs, a stochastic liquidity risk with an empirically validated nonlinear drift based on S&P 500 data, and two-factor stochastic volatility, the pricing problem takes the form of a five-dimensional highly nonlinear variational inequality. The proposed scheme decomposes this variational inequality at each time layer into an elliptic continuation equation approximated by a physics-informed neural network and a closed-form projection step enforcing the early-exercise constraint. A conditional $\ell^\infty$ stability bound is derived under stated resolvent and Lipschitz assumptions. The scheme is validated against finite-difference and least-squares Monte Carlo benchmarks, and both the option prices and the learned exercise boundary show close agreement with the reference solutions. Sensitivity analyses with respect to the key model parameters are also reported.

Complexity vs Empirical Score

• Math Complexity: 9.0/10
• Empirical Rigor: 7.0/10
• Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper presents a highly complex mathematical framework involving variational inequalities, operator splitting, and PINNs for a five-dimensional problem. It demonstrates strong empirical rigor through validation against benchmarks and an empirically validated nonlinear drift based on S&P 500 data. The deep learning-driven operator splitting for free boundary problems in this complex market setting is a novel contribution.

Research Flowchart

  flowchart TD
    A[Research Goal: Solve High-Dimensional Free Boundary Problems, e.g., American Option Pricing] --> B{Key Methodology: Deep Learning-Driven Operator Splitting Scheme};
    B --> C[Method Steps: Decompose Variational Inequality into Elliptic Continuation (PINN) & Projection Step];
    C --> D[Inputs: 5D Variational Inequality (Proportional Transaction Costs, Stochastic Liquidity Risk, Two-Factor Stochastic Volatility)];
    D --> E[Computational Processes: PINN for Elliptic Eq., Closed-Form Projection, Conditional L-infinity Stability Analysis];
    E --> F[Outcomes: Validated Option Prices & Learned Exercise Boundary (Close agreement with benchmarks), Sensitivity Analysis];

Authors: Saeid Safarveisi

Abstract

This paper presents a framework for pricing catastrophe (CAT) bonds in incomplete markets through Riemannian geometric optimization of financial entropy. We formulate pricing-measure selection as a maximum financial entropy problem on a finite-dimensional manifold of equivalent measures, parameterized by the non-hedgeable market prices of catastrophe frequency and severity risk, with discounting fixed by the default-free term structure. Since unconstrained entropy maximization recovers the physical measure, we incorporate observed secondary-market spreads, both pooled and date by date. For a compound Poisson-generalized Pareto loss model with Wang distortion, we derive closed-form expressions for financial entropy and the Riemannian metric. We show that CAT bond prices depend on frequency and severity premia primarily through the risk-neutral trigger intensity, making the two premia unidentifiable from market data alone; the entropy criterion resolves this ambiguity. Using a hypothetical earthquake CAT bond calibrated to quarterly spreads of Golden State Re II Ltd. (Series 2014-1), we obtain risk-neutral trigger intensities of 6.58 to 10.45 times the physical intensity (9.20 pooled and 18.26 at the end of 2015). The model-implied risk multiple aligns with the market spread-to-expected-loss ratio (8.87 versus 8.81 at issuance). The minimal-entropy measure attributes 53.5% to 63.7% of the log risk multiple to frequency risk, reflecting the selection principle rather than market identification. Finally, equating the coupon with the quoted spread understates the risk multiple by approximately half.

Complexity vs Empirical Score

• Math Complexity: 9.0/10
• Empirical Rigor: 7.0/10
• Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: This paper demonstrates a high level of mathematical sophistication, particularly with Riemannian geometry and financial entropy. It also exhibits strong empirical rigor through calibration to real market data and robust analysis of risk premia. The novelty lies in its unique approach to resolving identification issues in incomplete markets.

Research Flowchart

  flowchart TD
    A[Research Goal: Price CAT Bonds in Incomplete Markets & Unbundle Risk Premia] --> B{Methodology: Riemannian Geometric Optimization of Financial Entropy};
    B --> C{Model: Compound Poisson-Generalized Pareto Loss with Wang Distortion};
    C --> D[Inputs: Hypothetical Earthquake CAT Bond, Quarterly Spreads of Golden State Re II Ltd. (Series 2014-1)];
    D --> E[Computation: Derive Closed-Form Expressions for Entropy & Metric, Optimize for Risk-Neutral Trigger Intensity];
    E --> F[Outcome 1: Risk-neutral Trigger Intensities (6.58x to 10.45x physical; 9.20x pooled)];
    F --> G[Outcome 2: Model-implied risk multiple (8.87) aligns with market spread-to-expected-loss ratio (8.81)];
    G --> H[Outcome 3: Minimal-entropy measure attributes 53.5%-63.7% of log risk multiple to frequency risk];
    H --> I[Outcome 4: Equating coupon with quoted spread understates risk multiple by ~half];

Authors: Alexander Wu, Jingchao Huang

Abstract

Recurring cash flows are the backbone of cash-flow underwriting: they separate durable revenue from one-off deposits, expose standing obligations, and ultimately determine ability to pay. For small businesses, recurrence is both indispensable and hard to estimate, because much of their income arrives as ambiguous transfers, most payment streams are short, and schedules follow calendar conventions that frequency-domain methods treat as noise. We present CADENCE, Calendar-Aware Distributional Event Network for Cash-Flow Estimation, a compact causal Transformer that treats each transaction as a token and predicts the amount and a full, nonparametric timing distribution of every cash flow’s next payment. Inheriting transaction semantics from an external labeler and a frozen text encoder keeps CADENCE compact yet general across the long tail of small-business counterparties, while predicting a calibrated distribution rather than an opaque embedding makes its outputs interpretable enough to feed existing underwriting models and power credit decisions. Ablations confirm the role of each design choice, with counterparty semantics driving recall on short histories, calendar inputs driving timing accuracy, and no-event supervision driving precision. Against the production FFT baseline, CADENCE improves one-day timing accuracy by 93.3% and next-event F1 by 6.7%, while producing calibrated distributions that serve directly as a confidence gate for downstream decisions. In an end-to-end credit evaluation on production applications, replacing FFT recurrence features with CADENCE features raises KS by 4.8%, raises the approval rate at a fixed loss rate by 18.2%, and lowers the loss rate at a fixed approval rate by 13.1%.

Complexity vs Empirical Score

• Math Complexity: 6.0/10
• Empirical Rigor: 8.0/10
• Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper presents a sophisticated Transformer-based model for a complex financial forecasting problem, demonstrating strong empirical results against a production baseline. Its novelty lies in adapting language modeling techniques to business cash flow, with a clear focus on practical application and interpretability.

Research Flowchart

  flowchart TD
    A[Research Goal: Improve Small Business Cash-Flow Forecasting for Underwriting] --> B{Key Methodology: CADENCE Model Development};
    B --> C[Data/Inputs: Transaction Histories, External Labeler Semantics, Frozen Text Encoder, Calendar Features];
    C --> D{Computational Process: Causal Transformer (CADENCE) Training};
    D --> E[Outcomes: Predict Next Payment Amount & Nonparametric Timing Distribution];
    E --> F[Key Findings: Improved Timing Accuracy (93.3%), Next-Event F1 (6.7%), Calibrated Distributions, Increased KS (4.8%), Higher Approval (18.2%), Lower Loss Rate (13.1%)];

Authors: Minh Duc Duong, Alessio Moneta, Fabrizio Lillo, Caterina Giannetti, Tu Anh Doan Nguyen

Abstract

We study how contemporaneous equity-return transmission shapes systemic exposure among 51 listed European banks. Using a recursive non-Gaussian network model on volatility-standardized returns, we construct two market-capitalization-weighted measures based on direct and indirect transmission effects: the Systemic Impact Score (SIS), capturing a bank’s potential influence on other institutions, and the Systemic Vulnerability Score (SVS), capturing its exposure to disturbances originating elsewhere. The fitted network reveals distinct transmitting and receiving roles, highlighting differences between institutions with substantial outward influence and those with high incoming sensitivity. Receiving roles are generally more reproducible than transmitting roles across repeated network discovery, and selected connections remain informative about lower-tail dependence in subsequent observations. Volatility standardization substantially reduces detected dependence among recovered shocks, although the full-sample model does not satisfy the designated criterion for supporting shock independence. Alternative recursive orderings also yield comparable directional tail asymmetries, limiting the extent to which these patterns validate a particular ordering. The results provide complementary measures of model-implied systemic transmission and vulnerability, with explicit attention to uncertainty and identification limits

Complexity vs Empirical Score

• Math Complexity: 7.5/10
• Empirical Rigor: 8.0/10
• Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper employs advanced non-Gaussian network models (VAR-LiNGAM) for causal discovery in financial networks, demonstrating high mathematical sophistication. It applies these methods to real-world European banking data with careful consideration of statistical assumptions and robustness checks, indicating strong empirical rigor. The combination of these elements places it firmly in the ‘Holy Grail’ quadrant.

Research Flowchart

  flowchart TD
    A[Research Goal: Assess Contemporaneous Transmission & Systemic Risk in European Banking] --> B{Key Methodology: Recursive Non-Gaussian Network Model};
    B --> C[Data/Inputs: Equity Returns from 51 Listed European Banks];
    C --> D[Computational Process: Volatility Standardization & Network Construction];
    D --> E{Computational Process: Systemic Impact Score (SIS) & Systemic Vulnerability Score (SVS) Calculation};
    E --> F[Key Findings: Distinct Transmitting & Receiving Roles, Reproducible Receiving Roles, Informative Connections for Tail Dependence, Tail Asymmetries];

Authors: Ahzam Khan, Cavey Cool

Abstract

In utilization dependent interest rate models, adding liquidity can reduce a supplier’s total interest income even as the supplier’s balance increases. A supplier holding a fraction $α$ of final pool supply keeps income nondecreasing with a marginal deposit exactly when $α(1+η)\leq1$, where $η$ is the elasticity of the borrow rate with respect to utilization. When the condition fails, the supplier can instead raise its income by withdrawing, which pushes utilization further up. On two public Drift markets, a lender held more of the pool than the curve protected in at least one hour in nine on SOL and one in thirteen on USDC. A protocol turns this into a design constraint by choosing, at each utilization, the largest supplier share it intends to protect. That choice caps the elasticity, and integrating the cap describes every admissible curve. Protecting the same share at every utilization gives a power law boundary, and protecting every fresh depositor gives a hyperbolic one. On a multiple kink USDT market, compliant repairs can lower or raise supplier income, because the constraint governs how income responds to a deposit and not the price level.

Complexity vs Empirical Score

• Math Complexity: 7.5/10
• Empirical Rigor: 7.0/10
• Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper presents a novel and critical insight into pooled lending market design, backed by rigorous mathematical derivation and empirical validation on real-world data. The methodology is clearly explained, and the provision of Lean 4 formalizations and analysis code significantly enhances reproducibility.

Research Flowchart

  flowchart TD
    A[Research Goal: Understand Income Monotonicity in Pooled Lending] --> B{Key Problem: Can adding liquidity reduce supplier income?};
    B --> C[Methodology: Analyze supplier income dynamics with utilization-dependent rates];
    C --> D[Model Input: Supplier share α, Borrow Rate Elasticity η];
    D --> E[Analysis: Derive condition for non-decreasing income: α(1+η) ≤ 1];
    E --> F[Empirical Data: Drift Markets (SOL, USDC) & Multiple Kink USDT Market];
    F --> G[Computational Check: Identify hours where α(1+η) > 1];
    G --> H[Outcomes: Income monotonicity failure observed; Protocol design constraint derived; Admissible rate curves (power law, hyperbolic) described; Impact of repairs on USDT market];

Authors: Jonathan Chávez Casillas, Li Gan

Abstract

We study optimal quote placement for a market maker facing order flow modeled by a state-dependent Hawkes process. A block-based linear demand specification links market-order counts to execution volumes, providing the demand moments required by a continuous time discrete state dynamic programming formulation. Under a quadratic terminal inventory penalty, we characterize inventory-dependent quotes and derive backward recursions for the coefficients of a quadratic value function. Numerical experiments using LOBSTER data examine quote distributions, inventory exposure, and terminal wealth. The implemented strategy yields higher aggregate terminal wealth than fixed-spread benchmarks at one and three ticks, but lower wealth than the two-tick benchmark. These results illustrate the framework’s numerical implementation and distinguish terminal-wealth comparisons from evaluation under the inventory-penalized objective.

Complexity vs Empirical Score

• Math Complexity: 8.5/10
• Empirical Rigor: 6.0/10 (capped from 7.5: no out-of-sample test; see penalty flags)
• Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: This paper presents a highly mathematical framework for optimal order placement using state-dependent Hawkes processes, demonstrating strong empirical validation with LOBSTER data. The novelty lies in integrating sdHawkes processes into a stochastic optimal control framework for market making, which is a significant contribution. While the methodology is detailed, explicit code or data links for full reproducibility are not mentioned in the excerpt.

Research Flowchart

  flowchart TD
    A[Research Goal: Optimal Quote Placement for Market Maker] --> B{Methodology: State-Dependent Hawkes Process + Dynamic Programming};
    B --> C[Model Inputs: LOBSTER Data, Block-based Linear Demand, Quadratic Terminal Inventory Penalty];
    C --> D{Computational Process: Backward Recursions for Quadratic Value Function Coefficients};
    D --> E[Key Findings: Inventory-Dependent Quotes Characterized];
    E --> F{Outcomes: Higher Aggregate Terminal Wealth vs. 1 & 3-Tick Fixed Spreads};
    F --> G[Further Outcome: Lower Wealth vs. 2-Tick Fixed Spread, illustrating objective differences];

Authors: José Da Fonseca, Yajun Shi, Wenjun Zhang

Abstract

This paper introduces a model for infinite-server queues driven by a Cox process with stochastic intensity given by a squared Bessel (BESQ) process. Exploiting the affine structure of the underlying dynamics, we derive a closed-form expression for the joint moment-generating function (MGF) of the arrival intensity, the number of customers in the system, and the integral of the number of customers in the system over a time interval. This MGF can also be obtained at two (or more) dates, thereby allowing an analysis of the time evolution of infinite-server queues driven by a Cox–BESQ process. The availability of these explicit MGFs enables the development of risk measures for these variables, such as the joint moments of the arrival intensity, the number of customers in the system, and the integral of the number of customers in the system over a time interval, conditional on one of these variables exceeding a critical threshold. These risk measures can be determined at one or two dates. They rely essentially on the explicit form of the MGF of the state variables and on Fourier representations of the risk measures. Numerical experiments demonstrate the practical implementation of the proposed methodology and highlight the relevance of the theoretical results.

Complexity vs Empirical Score

• Math Complexity: 9.0/10
• Empirical Rigor: 4.0/10
• Quadrant: Lab Rats — theoretically deep, empirically untested

Why this score: The paper is highly theoretical, focusing on complex mathematical derivations for risk measures in a novel queueing model. While it includes numerical experiments, the empirical rigor in terms of real-world data application or extensive backtesting is limited. The novelty lies in the analytical tractability of the Cox-BESQ model for risk measures.

Research Flowchart

  flowchart TD
    A[Research Goal: Develop Risk Measures for Infinite-Server Queues] --> B{Model Definition: Infinite-Server Queues with Cox-BESQ Process}
    B --> C[Key Methodology: Exploit Affine Structure]
    C --> D{Derivation: Closed-Form Joint MGF for Key Variables}
    D --> E[Computational Process: Fourier Representations for Risk Measures]
    E --> F[Numerical Experiments: Practical Implementation & Validation]
    F --> G[Key Findings: Explicit Risk Measures (Joint Moments, Conditional on Thresholds) for Queue Variables]

Authors: Oskar Edén Wallberg, Alexander Karolin

Abstract

Portfolio construction has been studied extensively in both academia and industry. Its applications range from pension funds and hedge funds to private savings. The academic literature centers on methodology, commonly evaluated in settings that allow short selling, leverage, and frequent trading. Industry, on the other hand, operates under real constraints, but its data and methods are proprietary and largely unpublished. Neither is directly available to the private investor. This paper is an empirical study of established portfolio construction methods in the setting of the private investor. We consider a small investable universe of liquid US ETFs, where trading is monthly and carries a cost, and neither shorting nor leverage is permitted. Portfolio construction is posed as a convex optimization problem, its inputs are estimated from public data, and every parameter is interpretable. We denote such portfolios simple. We show that simple portfolios of as few as three ETFs deliver consistent and decent returns, without tuning to the backtest period. To find them, we have developed a framework, which we share publicly, able to backtest hundreds of portfolios on a standard laptop over a twenty year period or longer. When no return forecast is made, portfolios are constructed from risk alone, using methods such as risk parity and minimum volatility, built on a statistical factor model fitted with expectation maximization. When a forecast is made, we combine the Markowitz formulation with momentum signals and ridge regression. We report what works and what each component adds, weighing its gain against its complexity and the effort it demands of the investor. The findings are supported by numerical experiments over the past two decades, and the proposed portfolios outperform the standard 60/40 allocation.

Complexity vs Empirical Score

• Math Complexity: 6.5/10
• Empirical Rigor: 8.0/10
• Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: This paper demonstrates a strong blend of mathematical rigor in its portfolio optimization and risk modeling, coupled with extensive empirical validation through backtesting over two decades. The novelty lies in adapting established methods to the specific constraints and needs of private investors, providing a practical and reproducible framework.

Research Flowchart

  flowchart TD
    A[Research Goal: Deliver Consistent Decent Returns for Private Investors] --> B{Methodology: Empirical Study of Established Portfolio Construction Methods};
    B --> C{Inputs: Liquid US ETFs, Monthly Trading, No Shorting/Leverage, Public Data};
    C --> D{Computational Processes: Convex Optimization, Risk Parity, Min Volatility, Factor Model (EM), Markowitz + Momentum (Ridge Regression), Backtesting hundreds of portfolios (20+ years)};
    D --> E[Key Findings: Simple Portfolios (3+ ETFs) Deliver Consistent Decent Returns, Outperform 60/40, Framework Shared Publicly];

Authors: Haotong Fu, Xiyue Zhang, Jack Rose, Ding Sun, Erich Scheller, Eric Rose

Abstract

In healthcare equities, much of what moves a single name arrives as text: filings, trial and regulatory records, news and analyst commentary. We evaluate how screened non-market features and language-model representations improve market-based equity rankings for US healthcare firms over 2020-2025. Adding balance-sheet and attention features improves the resulting hybrid’s price- and liquidity-neutral rank information coefficient (IC) at one, three and twelve months. At twelve months, neutral IC rises from 0.063 to 0.093, a 48% increase, while raw IC rises from 0.289 to 0.298. Language models contribute most clearly at three months: among company-months with scientific-text coverage, a current embedding model raises the hybrid’s neutral IC from 0.046 to 0.056. This increment is nominally significant (t = 2.8; paired 95% bootstrap interval [+0.006, +0.015]) and remains exploratory. A larger post hoc gain occurs in months containing trial or regulatory records, consistent with text complementing prices and financial statements at that horizon. Across all text, pre-2020 and current representations yield small positive three-month increments in both neutral and raw IC but no established twelve-month gain. Much of their ranking performance survives within-company time shuffling, consistent with a persistent company component. Rater-cutoff comparisons find negligible pre-cutoff advantages for current models over the pre-2020 rater at one and three months that do not shrink after the cutoffs; twelve-month hindsight attribution remains inconclusive. At twelve months, balance-sheet signals lead after rallies and in volatile markets, while attention leads after declines. Retrospective state-dependent blending improves on a static blend without an established advantage over the full hybrid. Results apply to a survivor-only sector sample and remain unadjusted for the full exploratory search.

Complexity vs Empirical Score

• Math Complexity: 5.0/10
• Empirical Rigor: 8.0/10
• Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper presents a robust empirical evaluation of LLMs in equity ranking, utilizing detailed statistical methods and a clear backtesting framework. While the mathematical derivations are not explicitly shown in the excerpt, the methodology implies a solid quantitative foundation. The novelty lies in the specific application and comparative analysis of LLMs within a hybrid ranking system for healthcare equities.

Research Flowchart

  flowchart TD
    A[Research Goal: Evaluate LM Impact on Healthcare-Equity Rankings] --> B{Key Methodology Steps};
    B --> C[Data/Inputs: Screened Non-Market Features, LM Representations, Filings, Trial/Regulatory Records, News, Analyst Commentary, Balance-Sheet Data];
    C --> D[Computational Processes: Hybrid Model Development, Price- & Liquidity-Neutral Rank Information Coefficient (IC) Calculation, Embedding Models, State-Dependent Blending];
    D --> E{Key Findings/Outcomes};
    E --> F[Outcome 1: Balance-sheet & Attention Features improve hybrid's neutral IC (48% increase at 12m)];
    E --> G[Outcome 2: LMs contribute most at 3m (neutral IC from 0.046 to 0.056 for scientific-text coverage)];
    E --> H[Outcome 3: Larger post hoc gain in months with trial/regulatory records];
    E --> I[Outcome 4: Pre-2020 & current LMs yield small positive 3m IC increments, no established 12m gain];
    E --> J[Outcome 5: Balance-sheet signals lead after rallies/volatile markets; attention after declines];

Authors: Tiantian Wu, Peng Yang

Abstract

Quantitative investment requires alpha factors that are both predictive and diverse. However, existing factor generation methods, including formulaic mining and neural end-to-end models, are mostly based on a single-agent design, which limits exploration of the factor space and reduces adaptability to changing markets. Multi-agent approaches can broaden exploration, but agents trained under a shared objective often collapse to similar outputs, while existing diversity techniques either scale poorly with asset-universe size or address redundancy only after training. We propose a multi-agent sequential optimization (Det-MASO) framework for generating diversified, context-adaptive alpha factors. It introduces a differentiable log-determinant penalty on the inter-agent correlation matrix directly into the joint objective, discouraging correlated agent behavior during training with computational cost that scales linearly with the number of assets. We further show that this penalty prevents homogeneous equilibria and preserves factor diversity throughout optimization, while a context-conditioned scheduler dynamically adjusts agent emphasis across market regimes and Shapley interaction index analysis is used to characterize the cooperation structure that emerges among them. Experiments show that Det-MASO outperforms baselines, improving the information coefficient by 1.1%, the annualized return by 6.4%, and the Sharpe ratio by 0.33.

Complexity vs Empirical Score

• Math Complexity: 7.5/10
• Empirical Rigor: 6.0/10 (capped from 8.0: transaction costs not modelled; see penalty flags)
• Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper presents a highly novel multi-agent approach to factor generation with a sophisticated differentiable determinant penalty. It includes theoretical propositions and strong empirical results, indicating a robust methodology and significant contribution to quant finance. The experimental results with IC, annualized return, and Sharpe ratio improvements demonstrate strong empirical rigor.

Research Flowchart

  flowchart TD
    A[Research Goal: Generate Diversified, Context-Adaptive Alpha Factors] --> B{Problem: Single-agent limits exploration, multi-agent diversity issues};
    B --> C[Methodology: Det-MASO Framework];
    C --> D{Key Components:
        - Differentiable Log-Determinant Penalty
        - Context-Conditioned Scheduler
        - Shapley Interaction Index Analysis
    };
    D --> E[Computational Processes: Multi-Agent Sequential Optimization];
    E --> F[Outcomes:
        - Improved Information Coefficient (1.1%)
        - Increased Annualized Return (6.4%)
        - Higher Sharpe Ratio (0.33)
        - Preserved Factor Diversity
        - Adaptability to Market Regimes
    ];

Authors: Jean-François Chassagneux, Stéphane Crépey, Grégoire Loeper, Justin Ruelland

Abstract

We study the dependence on the terminal condition of solutions to a fully nonlinear parabolic Hamilton–Jacobi–Bellman (HJB) equation, when the terminal condition varies along a finite-dimensional family. This problem is motivated by a dual formulation of semimartingale optimal transport. The equation is posed as a terminal-value Cauchy problem on an unbounded spatial domain. Under structural ellipticity and regularity assumptions on the nonlinear operator, we recall local parabolic Hölder estimates and stability results for bounded uniformly continuous terminal data. We then establish continuous directional differentiability of the solution map with respect to the finite-dimensional parameter, and identify its derivative through the corresponding linearized Cauchy problem. Under an additional convexity assumption on the Hamiltonian and a non-degeneracy condition on the terminal data, we prove local strong convexity of the point-evaluated solution map. The analysis relies on Evans–Krylov theory for fully nonlinear parabolic equations and combines PDE arguments with probabilistic representations. Finally, we extend the differentiability and convexity results to cascades of Cauchy problems in which successive equations are coupled through terminal conditions at several prescribed times. This structure is motivated, for instance, by optimal transport problems with marginal constraints imposed at several maturities.

Complexity vs Empirical Score

• Math Complexity: 9.0/10
• Empirical Rigor: 1.0/10
• Quadrant: Lab Rats — theoretically deep, empirically untested

Why this score: This paper is highly theoretical, focusing on advanced PDE theory for HJB equations. While the mathematical derivations are rigorous and complex, there is no empirical testing or data validation presented. The novelty lies in extending differentiability and convexity results to cascades of Cauchy problems, which is relevant for multi-marginal optimal transport.

Research Flowchart

  flowchart TD
    A[Research Goal: Analyze Differentiability & Convexity of HJB Solution Maps w.r.t. Terminal Condition] --> B{Key Methodology: PDE Arguments & Probabilistic Representations};
    B --> C[Inputs: Fully Nonlinear Parabolic HJB Equation, Finite-Dimensional Terminal Condition Family];
    C --> D{Computational Processes: Evans-Krylov Theory, Local Parabolic Hölder Estimates, Linearized Cauchy Problem};
    D --> E[Outcome 1: Continuous Directional Differentiability of Solution Map & Derivative Identification];
    D --> F[Outcome 2: Local Strong Convexity of Solution Map under Additional Assumptions];
    F --> G[Outcome 3: Extension of Results to Cascades of Coupled Cauchy Problems];

Authors: Jiwon Jung, Kiseop Lee

Abstract

Prices cluster at round numbers, a pattern often interpreted as evidence that these levels act as price barriers. This pattern also admits a mechanical explanation: more liquidity rests there, and prices take longer to leave deeper queues. Limit order book data allow us to distinguish these mechanisms through a tick-grid framework in which local time counts departures and level-wise volatility reflects the departure hazard. Using data for five Nasdaq stocks throughout 2025, we find that round levels attract more liquidity but do not exhibit longer depth-adjusted durations. Intel and SoFi show directional effects, with departures biased away from round levels.

Complexity vs Empirical Score

• Math Complexity: 7.5/10
• Empirical Rigor: 8.0/10
• Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper presents a novel tick-grid framework using local time and duration analysis to investigate price clustering, offering a sophisticated statistical approach to a well-known market microstructure phenomenon. It combines advanced mathematical concepts with rigorous empirical testing on high-frequency LOB data, providing strong evidence against the ‘price barrier’ hypothesis for round numbers after adjusting for liquidity.

Research Flowchart

  flowchart TD
    A[Research Goal: Do Round Numbers Hold Prices?] --> B{Methodology: Tick-Grid Framework};
    B --> C[Inputs: LOB Data for 5 Nasdaq Stocks (2025)];
    C --> D{Computational Processes: Local Time, Level-wise Volatility, Departure Hazard};
    D --> E[Outcome 1: Round Levels Attract More Liquidity];
    D --> F[Outcome 2: No Longer Depth-Adjusted Durations at Round Levels];
    F --> G[Outcome 3: Intel/SoFi Show Directional Effects (Departures Biased Away from Round Levels)];

Authors: Matthew Brigida

Abstract

Total value locked (TVL) is widely used to assess decentralized finance ecosystems. Prior evidence finds that portfolios formed on TVL alone are spanned by the crypto market portfolio (Brigida 2025). This paper asks whether TVL becomes more informative when combined with measures of adoption, usage, development, protocol revenue, or social attention. I construct weekly value-weighted 2x3 portfolios that independently sort tokens on TVL-to-market-cap and one of seven ecosystem-activity ratios. Portfolio returns are computed from token prices using lagged-market-capitalization weights and evaluated against crypto market and size factors. Across the available 2022-2025 samples, none of the six on-chain double sorts rejects the benchmark after controlling for multiple testing. The same conclusion holds over 83 common weeks for which social-attention data are available. Several specifications reject at unadjusted levels, but none survives Holm adjustment. The social-attention rejection is only marginal under asymptotic inference and is insignificant under a moving-block bootstrap. Equal-weighted portfolios and formal parameter-stability tests lead to the same broad conclusion. The evidence therefore does not show that ecosystem activity reliably transforms TVL into an incremental return signal.

Complexity vs Empirical Score

• Math Complexity: 5.0/10
• Empirical Rigor: 6.0/10 (capped from 8.0: transaction costs not modelled; no out-of-sample test; see penalty flags)
• Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper employs rigorous statistical methods, including multiple testing adjustments and bootstrap methods, to evaluate portfolio returns. While the mathematical derivations are not explicitly shown, the statistical methodology is advanced and well-applied. The findings, though negative, are informative and contribute to the understanding of DeFi metrics.

Research Flowchart

  flowchart TD
    A[Research Goal: Does Ecosystem Activity Validate TVL?] --> B{Methodology: Construct 2x3 Portfolios};
    B --> C[Data: TVL-to-Market-Cap & 7 Ecosystem Activity Ratios (2022-2025)];
    C --> D{Computation: Weekly Value-Weighted Portfolio Returns};
    D --> E{Analysis: Factor Model Evaluation vs. Crypto Market & Size Factors};
    E --> F[Outcome: No Significant Incremental Return Signal After Multiple Testing Adjustment];

Authors: Hyoeun Lee

Abstract

Hawkes models of price changes in the consolidated US quote feed find a sub-millisecond kernel component and a branching ratio near criticality. We show that in large-tick stocks the first is the feed’s construction: half of NBBO changes are locks or depleted quotes, resolved within a millisecond, which the model reads as excitation. Defining the event as a mid-price level persisting for one millisecond removes the component. Raw fits put three times the clean kernel mass within one second, which a simulation reproduces. Since the total branching ratio trades off against the baseline, we report kernel mass on a fixed window.

Complexity vs Empirical Score

• Math Complexity: 7.5/10
• Empirical Rigor: 8.0/10
• Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: This paper demonstrates strong empirical rigor through its detailed data analysis, simulation studies, and robust statistical methodology, particularly in defining and testing event definitions. Its novelty lies in uncovering and quantifying the impact of consolidated feed construction on Hawkes model estimates, a previously unexamined aspect. The mathematical complexity is appropriate for the Hawkes process modeling and estimation techniques employed.

Research Flowchart

  flowchart TD
    A[Research Goal: Understand Price-Change Clustering in US Quote Feed] --> B{Methodology: Hawkes Models};
    B --> C[Data: Consolidated US Quote Feed Price Changes];
    C --> D{Computational Process: Fit Hawkes Model Parameters};
    D --> E{Analysis: Sub-millisecond Kernel & Branching Ratio Calculation};
    E --> F[Key Findings: Sub-millisecond kernel due to feed construction (locks/depleted quotes), branching ratio near criticality. Mid-price persistence event definition removes artifact. Raw fits overestimate kernel mass within 1 sec.];

Authors: Giorgos Iacovides, Oliver Hannibal, Wuyang Zhou, Danilo Mandic

Abstract

Recent advances in Generative AI have substantially improved financial sentiment analysis through post-trained financial large language models (LLMs). However, existing approaches remain confined to predicting the sentiment of each article in isolation and operate as black boxes, which limits contextual understanding and hinders deployment in high-stakes financial decision-making. To this end, we introduce FinLLM-KG, a model-agnostic framework that combines causal LLMs with a financial knowledge graph, in order to inject temporally grounded, company-level knowledge from SEC 10-K filings at inference time, without updating model weights. Within FinLLM-KG, context is supplied either as a natural-language company profile or as a small set of relevance-ranked factual triples. This enables sentiment predictions to account for broader firm-level information rather than the article alone, at almost zero computational cost and for any causal LLM. In addition, attributing each article-level sentiment score to individual sentences and to the retrieved knowledge-graph facts makes the evidence behind the prediction inspectable, thereby enhancing explainability and supporting human-in-the-loop review. Simulations demonstrate that FinLLM-KG consistently improves portfolio performance over the state-of-the-art baseline.

Complexity vs Empirical Score

• Math Complexity: 5.0/10
• Empirical Rigor: 6.0/10 (capped from 8.0: transaction costs not modelled; no out-of-sample test; see penalty flags)
• Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper presents a novel framework combining LLMs and KGs for financial sentiment, demonstrating improved portfolio performance through simulations. It emphasizes explainability and contextual understanding, addressing key limitations of existing black-box models. The empirical setup appears robust, though more details on the specific LLM and full backtesting methodology would enhance rigor.

Research Flowchart

  flowchart TD
    A[Research Goal: Improve Financial Sentiment Analysis for Algorithmic Trading with Explainability] --> B{Existing LLM Limitations};
    B --> C[Problem: Isolated Article Sentiment & Black Box Nature];
    C --> D[Proposed Solution: FinLLM-KG Framework];
    D --> E{Methodology: Combine Causal LLMs with Financial Knowledge Graph};
    E --> F[Data/Inputs: Financial Articles, SEC 10-K Filings, Company Profiles, Factual Triples];
    F --> G{Computational Process: Inference-time Knowledge Injection & Sentence/Fact Attribution};
    G --> H[Key Findings: Enhanced Contextual Understanding, Improved Explainability, Superior Portfolio Performance];

Authors: Pranay Anchuri

Abstract

Demand on blockchain-based marketplaces arrives in bursts and at a high rate. On these platforms the data are completely visible, which makes it feasible to compare the bursty demand against the usual load. The congestion manifests as an increased demand for gas, the primary resource a blockchain sells. Existing studies consider how to price the gas. The demand itself, and what it is made of during bursts, has not been measured. We measure it on Robinhood Chain, a layer-2 blockchain that quickly became a venue for trading tokenized equities and memecoins. We collected the data from every block of its first three active months and more than 50 million Uniswap swaps that occurred during bursty periods. We find that a burst trades about as many pools as a quiet period, mostly different ones, and trades them harder, on top of elevated demand that accumulated for hours. With the bursts aligned on their peaks, the median peak is 2.5 times the usual load. Both the rise and the fall of the demand are described by the sum of two exponential functions with different timescales, a slow one of hours and a fast one of minutes. Against a matched quiet period, a burst has 1.05 times the pools, 1.40 times the trades, and 1.21 times the senders. We use two models to explain the arrival of bursts: externally triggered cascades and self-excitation. The former fits the data relatively better. Inside a burst, the entry of each trader into a pool is followed for about 24 seconds by more entries than the existing traders predict. Whether this is because of imitation or because of a shared signal is left open.

Complexity vs Empirical Score

• Math Complexity: 6.5/10
• Empirical Rigor: 8.0/10
• Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: This paper demonstrates strong empirical rigor through extensive data collection and statistical analysis of blockchain demand bursts. Its novelty lies in measuring and modeling the shape and composition of these bursts, a previously unaddressed area. The mathematical models, while not overly complex, are appropriately applied to the empirical findings.

Research Flowchart

  flowchart TD
    A[Research Goal: Measure & Explain Blockchain Demand Bursts] --> B{Key Methodology: Quantify Demand Bursts};
    B --> C[Data/Inputs: Robinhood Chain Blocks (3 months), 50M Uniswap Swaps];
    C --> D{Computational Processes: Analyze Gas Demand, Uniswap Swaps, Trader Behavior};
    D --> E[Key Finding 1: Bursts trade more pools, harder, on top of accumulated demand];
    E --> F[Key Finding 2: Demand rise/fall described by two exponential functions (hours/minutes)];
    F --> G[Key Finding 3: Bursts have higher pools, trades, senders vs. quiet periods];
    G --> H[Key Finding 4: Externally triggered cascades fit better than self-excitation for burst arrival];
    H --> I[Key Finding 5: Trader entry into a pool followed by increased entries for ~24s (imitation/shared signal?)];

Authors: Dimitri Sotnikov, Adeline Viot

Abstract

We establish a martingality criterion for the Exponentially Fading Memory (EFM) signature volatility model, extending the recent result of Abi Jaber, Gassiat, and Sotnikov (2026). In the proposed framework, the volatility process is given by a linear functional of the EFM signature of a time-augmented Brownian motion, which yields a globally well-behaved, fully non-Markovian volatility model admitting a stationary distribution. The stationary EFM signature volatility model is also covered by allowing for a random initial value of the signature. Moreover, the volatility process can be adjusted to the initial forward variance curve. For this model, excluding trivial cases, we prove that the price process is a true martingale if and only if the truncation order is odd and a sign condition involving the correlation parameter holds.

Complexity vs Empirical Score

• Math Complexity: 9.0/10
• Empirical Rigor: 2.0/10
• Quadrant: Lab Rats — theoretically deep, empirically untested

Why this score: This paper is highly theoretical, focusing on advanced mathematical proofs for martingality in a novel volatility model. While the mathematical framework is sophisticated, there is no empirical validation or backtesting presented. The novelty lies in extending the EFM signature volatility model and establishing a new martingality criterion.

Research Flowchart

  flowchart TD
    A[Research Goal: Establish Martingality Criterion for EFM Signature Volatility Model] --> B{Key Methodology: Extend Abi Jaber, Gassiat, Sotnikov (2026) Result};
    B --> C[Model Formulation: Volatility as Linear Functional of EFM Signature of Time-Augmented Brownian Motion];
    C --> D{Model Characteristics: Globally Well-Behaved, Fully Non-Markovian, Stationary Distribution};
    D --> E[Inputs: Truncation Order, Correlation Parameter, Initial Forward Variance Curve];
    E --> F{Computational Proof: Martingality Condition Analysis};
    F --> G[Key Finding: Price Process is True Martingale IFF Truncation Order is Odd AND Specific Sign Condition Holds];

Authors: Naoufal Nifa

Abstract

We couple two sets of simulated market and credit paths by minimizing absolute temporal-correlation error relative to a prescribed target. Each market path receives exactly one credit path, preserving both empirical distributions of complete scenarios. A perfect-matching test identifies whether a uniform correlation-error tolerance is feasible and determines its smallest attainable value. Further assignments give exact empirical mean-loss and exceedance-probability bounds; quantile bounds follow by binary search over eligible loss thresholds. A synthetic bond-forward credit stress test illustrates the method. The experiments report matching precision, empirical score tails and exposure-grid error, and measure matching and mean-loss computation. The method is independent of the scenario generator and provides a reproducible way to examine dependence assumptions in existing risk engines. Tail measures concern default-probability-weighted credit scores. Accompanying reproducibility package (Python code, synthetic scenario libraries, permutations, and benchmarks) is attached in the ZIP archive.

Complexity vs Empirical Score

• Math Complexity: 7.5/10
• Empirical Rigor: 8.0/10
• Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper presents a mathematically sound approach to scenario matching using optimal transport and assignment problems. It demonstrates strong empirical rigor through a synthetic stress test, detailed numerical design, and analysis of matching precision and risk bounds. The methodology offers a novel way to examine dependence assumptions in existing risk engines, with a clear path to reproducibility.

Research Flowchart

  flowchart TD
    A[Research Goal: Reproducible Stress Test Dependence Assumptions] --> B{Key Methodology: Couple Market & Credit Paths};
    B -- Minimize --> C[Computational Process: Absolute Temporal-Correlation Error];
    C -- Inputs --> D(Data: Simulated Market Paths, Simulated Credit Paths);
    D -- Perfect Matching Test --> E{Outcomes: Smallest Attainable Uniform Correlation-Error Tolerance};
    E -- Further Assignments --> F[Outcomes: Exact Empirical Mean-Loss & Exceedance-Probability Bounds];
    F -- Illustration --> G(Synthetic Bond-Forward Credit Stress Test);
    G -- Analysis --> H[Key Findings: Matching Precision, Empirical Score Tails, Exposure-Grid Error, Method Performance];

Authors: Diogo Duarte, Lucas X. Ferreira, Emmanuel Gobet, Yuri F. Saporito

Abstract

We develop a no-arbitrage framework to recover implied fee rates from option-implied information and quantify fee compensation for liquidity providers in automated market maker (AMM) pools. Expected accumulated fees equal the current pool value minus the risk-neutral value of its terminal payoff, reducing forward-looking LP compensation to pricing the LP payoff. We apply the framework to Uniswap v2 and Uniswap v3 ETH–USDC pools using a stochastic-volatility jump-diffusion model calibrated to Deribit options. Implied fee rates are positive, time varying, downward sloping in maturity, and highly sensitive to range design, with tighter near-the-money ranges earning larger implied returns.

Complexity vs Empirical Score

• Math Complexity: 8.5/10
• Empirical Rigor: 7.5/10
• Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper presents a sophisticated no-arbitrage framework using stochastic-volatility jump-diffusion models, demonstrating high mathematical complexity. It applies this framework to real-world Uniswap data and Deribit options, showing strong empirical rigor. The novel approach of recovering implied fee rates from option-implied information is a significant contribution to DeFi research.

Research Flowchart

  flowchart TD
    A[Research Goal: Quantify LP Compensation & Implied Fee Rates] --> B{Methodology: No-Arbitrage Framework};
    B --> C[Data/Inputs: Uniswap v2/v3 ETH-USDC Pools, Deribit Options];
    C --> D{Computational Process: Calibrate Stochastic Volatility Jump-Diffusion Model};
    D --> E[Outcome 1: Implied Fee Rates are Positive, Time Varying, Downward Sloping];
    D --> F[Outcome 2: Fee Rates Highly Sensitive to Range Design (tighter ranges = larger returns)];
    E & F --> G[Conclusion: Developed framework for pricing LP payoff & assessing compensation];

Authors: Pedram Ebadollahyvahed, Nneka Umeorah

Abstract

Cryptocurrencies trade continuously while major traditional markets close over the weekend. We examine whether weekend cryptocurrency volatility predicts Monday variance in equity, commodity and fixed-income markets. Using daily range-based variance measures for Bitcoin, Ether, XRP, the S&P 500, gold and a US Treasury ETF over 2017-2026, we augment heterogeneous autoregressive models with a common weekend cryptocurrency factor. Predictive content is weak before 2024 but significant thereafter. In the post-2024 sample, a one-standard-deviation increase in the factor is associated with 27-54% higher fitted Monday variance. Adding the factor reduces out-of-sample QLIKE losses by 11-22%, although statistical evidence varies across markets. Friday cryptocurrency volatility does not reproduce the result, and coin-specific variation adds no significant predictive information beyond the common factor. An endogenous procedure estimates a break in late 2023, during growing anticipation of US spot-Bitcoin ETF approval. However, ETF-flow interactions provide no evidence of amplification through greater measured Monday flow activity, leaving the mechanism unresolved. The findings suggest that calendar-specific volatility measures reveal predictive information that aggregate connectedness measures may obscure.

Complexity vs Empirical Score

• Math Complexity: 6.0/10
• Empirical Rigor: 8.0/10
• Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper employs advanced econometric models (HAR, structural break analysis) and robust statistical tests (QLIKE, Diebold-Mariano) with a clear empirical focus. The novelty lies in exploiting the calendar mismatch for predictive insights, particularly the emergence of the relationship post-2023.

Research Flowchart

  flowchart TD
    A[Research Goal: Do Weekend Crypto Volatility Predict Monday Market Risk?] --> B{Methodology: Augment HAR Models with Common Weekend Crypto Factor};
    B --> C[Data Inputs: Daily Range-Based Variance (2017-2026) for BTC, ETH, XRP, S&P 500, Gold, US Treasury ETF];
    C --> D{Computational Process: Model Estimation & Predictive Power Assessment (In-sample R2, Out-of-sample QLIKE, Endogenous Break Analysis)};
    D --> E[Finding 1: Weak Predictive Content Pre-2024, Significant Post-2024 (27-54% higher fitted Monday variance)];
    E --> F[Finding 2: Out-of-sample QLIKE losses reduced by 11-22%; Friday crypto volatility & coin-specific variation are not predictive];
    F --> G[Finding 3: Endogenous Break in Late 2023 (US spot-Bitcoin ETF anticipation), but no evidence of amplification via ETF flows];
    G --> H[Outcome: Calendar-specific volatility reveals predictive information that aggregate connectedness measures may obscure].

Authors: Matin Tavakoli, Seyedreza Tavakoli, Faezeh Salmani, Amir Amini, Mohammad Hosein Rohban, Gholamali Aminian, Ali Habibnia Rohban

Abstract

Financial reinforcement learning often translates an investment objective into an additive risk-adjusted reward. An agent can optimise that reward accurately yet implement a different financial preference. We develop a decision audit for terminal mean-variance equilibrium that separates reward misspecification from learning error. In a finite exogenous-state market, we bound the gain from changing one action using actor, reward and continuation-critic residuals. The certificate ignores state-constant critic offsets, accommodates convex action constraints, and has an independent-sample extension that does not require the equilibrium policy. A continuation-aware actor-critic method estimates the future-policy correction from samples. Across nine controlled markets and twenty seeds, it reduces policy distance by and one-step deviation gain by relative to a matched myopic-reward network. A separate PPO comparison tests the same reward intervention within a standard optimiser. Historical tests cross seven covariance estimators and 10, 30 and 49 US industry portfolios over 432 months: joint reward calibration preserves each fitted mean-variance target, but does not improve realised investment performance. Together, these results provide a way to quantify whether a learned decision implements its stated financial objective, separately from how well that objective performs in a backtest.

Complexity vs Empirical Score

• Math Complexity: 8.0/10
• Empirical Rigor: 7.0/10
• Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper presents a mathematically dense framework for certifying reward alignment in RL, involving complex derivations and theoretical bounds. It supports this with rigorous empirical testing across controlled markets and historical data, demonstrating a strong blend of theoretical depth and practical validation. The novelty lies in the decision audit framework and continuation-aware learning for reward alignment.

Research Flowchart

  flowchart TD
    A[Research Goal: Certify Reward Alignment in Portfolio RL] --> B{Methodology: Decision Audit for Mean-Variance Equilibrium};
    B --> C[Inputs: Finite Exogenous-State Market, Convex Action Constraints];
    C --> D{Computational Process: Actor, Reward, & Continuation-Critic Residuals};
    D --> E[Computational Process: Continuation-Aware Actor-Critic (Future-Policy Correction)];
    E --> F[Validation 1: Controlled Markets (9 markets, 20 seeds)];
    E --> G[Validation 2: Historical Tests (7 Covariance Estimators, US Industry Portfolios, 432 months)];
    F & G --> H[Key Findings: Quantify Reward Misspecification vs. Learning Error; Improved Policy Distance & One-Step Deviation Gain (Continuation-Aware AC); Joint Reward Calibration Preserves Targets, but No Improved Realized Performance];

Authors: Alexandre Roch, Simone Scotti

Abstract

We study an optimal dividend policy problem under ratcheting and bankruptcy constraints. Firms face a tradeoff between the negative signal of a dividend reduction and a potential future bankruptcy. In contrast to previous works, the dividend rate is allowed to decrease after incurring a negative impact on equity valuation in accordance with a large empirical literature on the topic. We provide an explicit necessary and sufficient condition for the finiteness of the value function and prove the differentiability of the value function, and its characterization via viscosity solutions and the so-called supercontact property. The value function is given by the solution of a two-dimensional variational inequality combining optimal stopping, singular and impulse controls. We show that the problem is equivalent to a one-dimensional setup simplifying both mathematical and financial analysis. Finally, we provide a numerical study of the model and a sensitivity analysis with respect to the model parameters. We show the relevance of over-reaction to financial news on the manager’s optimal policy.

Complexity vs Empirical Score

• Math Complexity: 9.0/10
• Empirical Rigor: 4.0/10
• Quadrant: Lab Rats — theoretically deep, empirically untested

Why this score: This paper presents a highly complex mathematical framework for optimal dividend policies, introducing novel elements like dividend resets with equity valuation impacts. While it includes a numerical study and sensitivity analysis, the empirical rigor is limited to theoretical simulations rather than real-world data validation or backtesting. The overall contribution is strong due to its innovative theoretical approach.

Research Flowchart

  flowchart TD
    A[Research Goal: Optimal Ratcheting Dividend Policy with Resets] --> B{Key Methodology: Variational Inequality & Viscosity Solutions};
    B --> C[Data/Inputs: Equity Valuation, Dividend Rates, Bankruptcy Constraints];
    C --> D{Computational Process: Two-dimensional Variational Inequality Solved by Equivalent One-dimensional Setup};
    D --> E[Key Finding 1: Explicit Necessary & Sufficient Condition for Value Function Finiteness];
    D --> F[Key Finding 2: Differentiability of Value Function, Characterized by Supercontact Property];
    D --> G[Key Finding 3: Relevance of Over-reaction to Financial News on Manager's Policy];
    E & F & G --> H[Outcome: Numerical Study & Sensitivity Analysis of Optimal Policy];

Authors: Deepak Bastola, Eli Prushansky

Abstract

Compiling implied volatility (IV) surfaces and forecasting derivative risk from high-frequency equity and options feeds requires addressing contract heterogeneity, schema shifts, volatility clustering, and non-stationary trading activity. Using a split-adjusted NVIDIA (NVDA) equity and options panel (2018-2025), this study integrates finite-population survey sampling, tempered inverse-probability sequence learning, and causal activity clocks. Four findings emerge. First, cross-sectional surface accuracy depends on matching the sampling design to the target functional: plug-in stratified allocations using lagged cell dispersion improve equal-weighted surface reconstruction with full grid coverage, whereas liquidity-weighted probability proportional to size (PPS) sampling excels on liquidity-weighted targets by construction yet inflates error when misapplied. Second, in longitudinal LSTM sequence training, PPS and Neyman designs retain more tail-volatility windows than simple random sampling (SRS); all three designs are equivalent to full-census training within chosen margins while cutting training time, though incremental forecasting gains of informative selection over SRS are not established. Third, out-of-sample evaluation shows a predictability contrast: short-horizon prices are not forecastable beyond a random walk with drift among tested models, whereas forward IV shows descriptive mean-reversion skill concentrated in a few post-shock reversal sessions (insignificant under session-clustered tests), with sequence models not beating a linear HAR-IV benchmark and feature pruning driving stability. Fourth, across a 6.3x start-to-end (nearly twentyfold trough-to-peak) price expansion, rolling volume and dollar clocks stabilize bar rates and compress intraday return kurtosis beyond diurnal scaling.

Complexity vs Empirical Score

• Math Complexity: 8.5/10
• Empirical Rigor: 9.0/10
• Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: This paper demonstrates a high level of mathematical sophistication in integrating various statistical and machine learning techniques. Its empirical rigor is exceptional, with detailed data analysis, robust statistical testing, and clear out-of-sample evaluations. The novelty lies in the unique combination of survey sampling, sequence learning, and activity clocks for volatility forecasting.

Research Flowchart

  flowchart TD
    A[Research Goal: Compile & Forecast Implied Volatility (IV) Surfaces & Derivative Risk] --> B{Key Methodologies};
    B --> C[Data: Split-Adjusted NVDA Equity & Options Panel (2018-2025)];
    B --> D[Methodology: Finite-Population Survey Sampling (Stratified, PPS, SRS)];
    B --> E[Methodology: Tempered Inverse-Probability Sequence Learning (LSTM)];
    B --> F[Methodology: Causal Activity Clocks (Rolling Volume/Dollar Clocks)];
    C & D & E & F --> G[Computational Processes: Surface Reconstruction, LSTM Training, Out-of-Sample Evaluation, Feature Pruning];
    G --> H[Findings: 1. Sampling design impact on surface accuracy; 2. Sampling benefits in LSTM training; 3. Predictability contrast (prices vs. forward IV); 4. Activity clock stabilization.];

Authors: Derun Zhong, Guolian Wang, Shiqing Wang, Hongjie Feng

Abstract

Recent advances in financial machine learning have prompted growing interest in applying deep neural networks to convertible bond valuation and forecasting. However, existing studies often focus on in-sample valuation fit or cross-sectional return rankings, leaving out-of-sample next-day price forecasting insufficiently benchmarked against persistence. This paper provides a comprehensive empirical reality check on cross-bond panel learning for Chinese convertible bonds. Using a daily panel of 73,806 observations across 100 A-share convertible bonds from 2022 to 2025, we evaluate whether cross-bond neural architectures reliably outperform the random walk under strict temporal splitting, boundary purging, fold-specific preprocessing, and dependence-aware inference. On the 29,514-observation holdout, the random walk achieves an RMSE of 2.3577 CNY. Five-seed ensembles of small and large price-change MLPs yield RMSEs of 2.3619 and 2.3645 CNY, with negative out-of-sample R^2 values (-0.0035 and -0.0057) and target-date clustered HAC p-values of 0.453 and 0.248. Ten-day circular block-bootstrap intervals span zero, while direct price-level training degrades performance substantially. Although cross-bond pooling provides a variance-stabilizing shrinkage effect over single-bond estimation (lowering test RMSE for 79 of 90 individual bonds), it fails to produce an aggregate forecast advantage over persistence. Error decomposition reveals that failure is not uniform: five highly volatile, clause-sensitive bonds account for 74.4% of the model’s total excess squared error. Without contemporaneous equity market information or explicit contractual event conditioning, deep panel architectures do not deliver incremental next-day predictability over persistence.

Complexity vs Empirical Score

• Math Complexity: 6.0/10
• Empirical Rigor: 9.0/10
• Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: This paper demonstrates exceptional empirical rigor in its evaluation of deep learning models for convertible bond forecasting, employing strict temporal hygiene and robust statistical inference. While the mathematical complexity of the neural networks themselves is moderate, the rigorous econometric safeguards and detailed error decomposition contribute significantly to its overall strength and novelty in challenging existing assumptions.

Research Flowchart

  flowchart TD
    A[Research Goal: Next-Day Price Forecasting of Chinese CBs vs. Persistence] --> B(Methodology: Cross-Bond Panel Learning with Deep Neural Networks)
    B --> C{Data: Daily Panel of 100 A-share CBs (2022-2025, 73,806 obs)}
    C --> D[Computational Processes: Strict Temporal Split, Boundary Purging, Fold-Specific Preprocessing, Dependence-Aware Inference]
    D --> E(Evaluation: RMSE, Out-of-Sample R^2, HAC p-values, Block-Bootstrap Intervals)
    E --> F[Key Finding 1: MLPs Fail to Outperform Random Walk (RMSEs 2.36 vs 2.3577, neg R^2)]
    F --> G[Key Finding 2: Cross-Bond Pooling Reduces Individual Bond RMSE, but No Aggregate Advantage]
    G --> H[Key Finding 3: Failure Concentrated in 5 Volatile, Clause-Sensitive Bonds (74.4% of Excess Error)]
    H --> I(Conclusion: Deep Panel Architectures Lack Incremental Predictability Without External Info/Event Conditioning)

Authors: Lilian Hu, Ruodu Wang

Abstract

We provide axiomatic characterizations of Jensen-gap functionals, which have the form $ρ(X)=\mathbb{E}[f(X)]-f(\mathbb{E}[X])$ for some real-valued function $f$ called the generator. This class includes important examples such as the excess growth rate, the utility premium, the Kullback-Leibler (KL) divergence, and the variance, and is also closely related to measures of variability. We analyze various forms of convexity for Jensen gaps. To characterize Jensen gaps, we formulate conditional versions of law-invariant functionals, which are needed to state the decomposition principle. On an atomless probability space, we show that law invariance together with a binary decomposition principle is sufficient to characterize Jensen gaps on simple random variables. Under suitable continuity conditions, the characterization extends to broader classes, including compactly supported random variables and integrable random variables. We then identify the generator $f$ by imposing either of two natural invariance conditions: scaling invariance or translation invariance. The former identifies logarithmic, $x\log x$, and power generators, whereas the latter identifies exponential and quadratic generators.

Complexity vs Empirical Score

• Math Complexity: 9.0/10
• Empirical Rigor: 2.0/10
• Quadrant: Lab Rats — theoretically deep, empirically untested

Why this score: This paper is highly theoretical, focusing on axiomatic characterizations and mathematical proofs of Jensen-gap functionals. It lacks empirical validation or backtesting, placing it firmly in the ‘Lab Rats’ quadrant. The novelty lies in its comprehensive characterization of Jensen gaps and the decomposition principle.

Research Flowchart

  flowchart TD
    A[Research Goal: Axiomatic Characterization of Jensen-Gap Functionals] --> B{Key Methodology: Conditional Law Invariance & Decomposition Principle};
    B --> C[Data/Inputs: Simple, Compactly Supported, Integrable Random Variables];
    C --> D{Computational Process: Analyze Convexity Forms & Identify Generators};
    D --> E[Outcome 1: Characterization of Jensen Gaps on Atomless Spaces];
    D --> F[Outcome 2: Identification of Generators via Scaling/Translation Invariance (log, power, exp, quadratic)];
    E & F --> G[Key Findings: Jensen gaps are characterized by law invariance and decomposition; specific generators identified by invariance conditions];

Authors: Emma Kroell, Mogens Steffensen

Abstract

We study an investor’s optimal portfolio choice problem when they account for the impact of a possible future green transition, whose timing is uncertain. We model the green transition as a discontinuity in the stock price processes, driven by a single jump process with constant intensity, in which ‘green’ stocks increase in value and ‘brown’ stocks decrease in value. The investor is ambiguous about the timing of the green transition and designs an investment strategy which is robust to this risk. The investor has constant relative risk aversion preferences over both wealth and model ambiguity, which introduce a multiplicative ambiguity penalty for power utility and an additive ambiguity penalty for log utility. We separate the problem into pre- and post-transition cases, derive the solution, and provide a verification theorem. We further show that the log utility, ambiguity-neutral, and worst-case problems are obtained as special cases of our general problem. In a case study with one green and one brown stock, we show that when the stock prices are positively correlated, an investor who is aware of transition risk increases their investment in the green stock and reduces their investment in the brown stock, compared to when they ignore the possibility of a transition. We illustrate the difference between an ambiguity-averse and a risk-averse investor, and show how an ambiguity-averse investor re-allocates their portfolio as their view on the probability of a green transition changes.

Complexity vs Empirical Score

• Math Complexity: 8.5/10
• Empirical Rigor: 3.0/10
• Quadrant: Lab Rats — theoretically deep, empirically untested

Why this score: The paper presents a highly mathematical framework for optimal portfolio choice under green transition timing risk, incorporating ambiguity aversion. While the theoretical derivations are rigorous and novel, the empirical validation is limited to a single case study without extensive backtesting or real-world data analysis. This places it firmly in the ‘Lab Rats’ quadrant, excelling in theoretical development but needing more empirical grounding.

Research Flowchart

  flowchart TD
    A[Research Goal: Optimal Portfolio Choice under Green Transition Timing Risk] --> B{Key Methodology: Jump Process Model with Uncertain Timing & Ambiguity Aversion};
    B --> C{Inputs: Green/Brown Stock Prices, Jump Intensity, Investor Preferences (CRRA, Ambiguity Aversion)};
    C --> D[Computational Processes: Separate Pre/Post-Transition Cases, Derive Solution, Verification Theorem];
    D --> E[Key Findings: Increased Investment in Green Stocks, Reduced Brown Stocks under Transition Risk];
    E --> F[Further Findings: Differentiation of Ambiguity-Averse vs. Risk-Averse Investor Allocation; Log Utility & Worst-Case as Special Cases];

1. Arithmetic Models for Commodity Futures, Vanilla, and Calendar Spread Options

Holy Grail · Math 8.0/10 · Rigor 9.0/10 · Volatility, Commodities & Energy

We develop an arithmetic multi-factor stochastic volatility model for commodity futures and options. The arithmetic specification yields a low-dimensional Markovian representation of the futures curve and semi-closed-form pricing formulas for vanilla and basket options, including calendar spread opt…

2. A game-theoretic analysis of the market microstructure in decentralized exchanges with concentrated liquidity

Holy Grail · Math 9.0/10 · Rigor 8.0/10 · Crypto & DeFi, Market Microstructure

We study strategic liquidity provision in decentralized exchanges with concentrated liquidity as a mean-field game. Liquidity providers choose the width of their price range and when to rebalance it, pay a transaction cost at each rebalancing, and share the fees collected at each tick in proportion…

3. Arbitrage-Free Implied Volatility Surface Reconstruction via Additive Martingales

Holy Grail · Math 8.0/10 · Rigor 8.0/10 · Options & Derivatives, Volatility

This work focuses on reconstructing a continuous implied volatility (IV) surface from a small number of option quotes, a fundamental task for option pricing and risk management. We propose Additive Martingale (AddMart), a few-shot IV surface reconstruction framework that satisfies static no-arbitrag…

4. Teacher Supervision, Fixed Representations and Calibration: Exact Risk Calculations for Candidate-Scoring Decision Models

Holy Grail · Math 8.5/10 · Rigor 7.5/10

Candidate-scoring decision models return a probability distribution over a presented set of answers. Public documentation of one such system, Jev, and of an open counterpart, JevAny, describes three training ingredients that determine what these probabilities estimate: soft targets from larger model…

5. SHADE: Shrinkage via Halted Alpha-stable Denoising for Heavy-Tailed Probabilistic Time Series Forecasting

Holy Grail · Math 8.5/10 · Rigor 7.5/10 · Machine Learning

Time-series observations are particularly susceptible to noise, and many real-world series have low signal-to-noise ratios (SNRs) together with heavy-tailed and time-varying disturbances. When Gaussian diffusion models are used for forecasting, reversing all the way to the zero-noise endpoint may yi…

6. Understanding the Inner Workings of a State-Space Foundation Model for Limit Order Books

Holy Grail · Math 7.5/10 · Rigor 8.0/10 · Market Microstructure

Autoregressive foundation models offer state-of-the-art capabilities to simulate the market by generating the message flow of limit order books (LOBs). Yet, without understanding their inner workings, research directions are hard to narrow and practical use in real-world trading environments remains…

7. Cross-product Modelling in Continuous Intraday Electricity Markets via Hawkes Processes

Holy Grail · Math 7.5/10 · Rigor 8.0/10 · Market Microstructure, Commodities & Energy

As renewable energy integration accelerates, intraday continuous (IDC) electricity markets have become essential for short-term balancing. However, the simultaneous trading of overlapping contracts with varying granularities creates complex, hidden liquidity dependencies that are critical to market…

8. Amortized Neural Filtering for Calibrated Posteriors in Statistical Arbitrage

Holy Grail · Math 7.5/10 · Rigor 8.0/10 · Statistical Arbitrage

Statistical arbitrage trades on an estimate of the state of a residual spread, and conventional practice discards the uncertainty of that estimate. We keep the uncertainty and use it in the trading decision. A Neural Filter Estimator (NFE) is trained on simulated paths of a mean-reverting process wh…

9. Dynamics of the Liquidity Profile in Concentrated Liquidity Provision: An SPIDE Approach

Holy Grail · Math 9.0/10 · Rigor 7.0/10 · Crypto & DeFi, Market Microstructure

We model the dynamics of the liquidity profile in Uniswap~V3 and other concentrated-liquidity automated market makers with a stochastic partial integro-differential equation (SPIDE). The profile, the aggregate of all provider positions across tick space, evolves under a degenerate Sturm–Liouville o…

10. Conditional Weight of Evidence for Gradient-Boosted Credit Models: Estimation, Additivity and Its Relation to Adverse Action Reasons

Holy Grail · Math 7.5/10 · Rigor 8.0/10

Traditional weight of evidence assigns a credit characteristic a value based on its association with default across a population. Credit analysis also requires an assessment of that characteristic in the context of an individual applicant’s record. We study conditional weight of evidence for this pu…

Browse everything new in the flowchart library, or explore by area on the topic hubs.

Authors: Rischan Mafrur, Khadijah, Gun Gun Febrianza

Abstract

Security-token standards such as ERC-3643 decide who may hold a token, not what stands behind it. For a sukuk this gap matters: whether a certificate may trade at a negotiated price depends on what its portfolio contains, and a token that checks only investor identity keeps trading after every non-debt asset has been sold. We ask whether this tradability condition can be enforced on-chain, and at what cost. We encode the tradability conditions of the 2023 resolution of the Shariah Advisory Council of the Securities Commission Malaysia on Sukuk Wakalah bi al-Istithmar as issuance and transfer policies that read EIP-712 claims, signed by separate institutional roles, about each asset’s rights, valuation, encumbrance and Shariah classification and about the allocation of proceeds. Across 26 scenarios that we constructed on a synthetic instrument, an identity-only baseline wrongly allowed 13 operations that the encoded policy denies, and asset evidence without the Shariah rule still allowed 6. The full policy matched every expected outcome and added 35,834 gas (71.0%) per transfer, independent of portfolio size. Our evaluation also exposes a trade-off that the constructed scenarios had missed. Checking the continuing condition through a single nominated asset keeps the cost of a transfer constant but can block trades the ruling allows, and a bounded fallback scan narrows this gap at about 12,050 gas per scanned asset. Our results indicate that Shariah tradability conditions can be enforced on-chain against attributable asset evidence, and that the way a ruling is encoded shapes which lawful trades a token permits. GitHub repo: https://github.com/rischanlab/tokenized_sukuk

Complexity vs Empirical Score

• Math Complexity: 5.0/10
• Empirical Rigor: 7.0/10
• Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper presents a novel approach to enforcing Shariah compliance on-chain for tokenized sukuk, combining smart contract logic with verifiable off-chain attestations. Its empirical rigor is demonstrated through scenario-based evaluation and gas cost analysis, while the mathematical complexity lies in the logical encoding of compliance rules rather than advanced financial models.

Research Flowchart

  flowchart TD
    A[Research Goal: Enforce Sukuk Tradability On-Chain?] --> B{Methodology: Encode Shariah Rules as Policies};
    B --> C[Inputs: EIP-712 Claims (Asset Rights, Valuation, Encumbrance, Shariah Class)];
    C --> D{Computational Process: On-Chain Issuance & Transfer Policy Enforcement};
    D --> E[Scenarios: 26 on Synthetic Sukuk];
    E --> F[Outcomes: Full Policy Correctness, Gas Cost (35,834 per transfer), Trade-off (Single Asset vs. Fallback Scan)];
    F --> G[Finding: Shariah Tradability Enforceable On-Chain, Encoding Shapes Permitted Trades];

Authors: Li Guo, Frank Weikai Li, Ding Qi, Lijian Sun

Abstract

We develop a new order-level proxy for high-frequency trading (HFT) activity based on precise order placement and study its effects on market quality. With quotes updating multiple times per second, millisecond-level order-book access makes high-frequency traders more likely to submit active exact-match orders, defined as limit orders priced exactly at the prevailing best opposite-side quote. Using Shenzhen Stock Exchange data, we validate the proxy: active exact-match orders respond faster to quote updates, cancel more rapidly, concentrate when depth imbalance is favorable and spreads are narrow, and predict mid-quote returns from 100 milliseconds through ten minutes. A difference-in-differences design exploiting cross-channel variation in order dissemination latency at the market open shows that lower latency increases active exact-match shares. This variation yields plausibly causal evidence that latency-induced HFT raises short-horizon price variation, wide-spread incidence, and adverse-selection exposure, while realized spreads decline and pricing error falls modestly. Unlike developed markets, China’s T+1 constraint, absence of maker-taker rebates, and limits on rapid quote revision tilt speed advantages toward directional liquidity taking, highlighting how market design shapes the channel through which speed affects market quality.

Complexity vs Empirical Score

• Math Complexity: 6.0/10
• Empirical Rigor: 6.0/10 (capped from 8.0: no out-of-sample test; see penalty flags)
• Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper presents a novel empirical proxy for HFT activity and rigorously validates it using high-frequency data. The difference-in-differences design provides a strong causal identification strategy, contributing significantly to the understanding of HFT in a unique market setting.

Research Flowchart

  flowchart TD
    A[Research Goal: Track HFT in China and Effects on Market Quality] --> B{Develop New Order-Level HFT Proxy};
    B --> C[Data: Shenzhen Stock Exchange Order Book Data];
    C --> D{Validate HFT Proxy: Exact-Match Orders & Response to Quote Updates};
    D --> E{Causal Inference: Difference-in-Differences using Order Dissemination Latency};
    E --> F[Outcomes: HFT Raises Price Variation, Wide Spreads, Adverse Selection; Lowers Realized Spreads, Pricing Error];
    F --> G[Context: China's Market Design (T+1, No Maker-Taker) Favors Directional Liquidity Taking];

Authors: Dung Cao, Palaash Gang, Arjun Veluri

Abstract

We audit the scoring pipeline of FLARE, the financial task suite behind PIXIU and FinBen, and find that in a decisive fraction of cells the score is a property of the answer-extraction rule, not the model. The clearest evidence is already published: of the 34 released Headlines scores, nine equal to four decimals what the benchmark’s own code returns when no output is byte-equal to Yes, and only four clear the per-stratum majority floor. FLARE credits the first entry of the dataset card’s choices list found anywhere in a generation; the lm-evaluation-harness it forks takes the first match by position; on identical generations the two differ by up to 0.4216 accuracy, and no leaderboard states which rule produced its numbers. An unspecified rule leaves the score partially identified, and because the coarsening is multiplicity rather than missingness the interval’s width is counted rather than bounded. Across a 204-cell factorial and a 46-cell diagnostic, 10 cells are certified invariant to the rule, 4 only because almost nothing parses, so certification is paired with coverage. On Financial PhraseBank four of seven models fall below the 0.5948 majority floor under one shipped rule and clear it under the other; on three of four substring-scored tasks no open-weight model we test is distinguishable from a degenerate predictor. Reading a FLARE score needs three quantities it does not report – ambiguity rate, coverage, floor – and floorbench computes all three from generations already produced.

Complexity vs Empirical Score

• Math Complexity: 4.0/10
• Empirical Rigor: 6.0/10 (capped from 8.0: no out-of-sample test; see penalty flags)
• Quadrant: Street Traders — practical and empirical, lighter on theory

Why this score: This paper rigorously audits a critical component of LLM financial benchmarks, revealing significant issues with answer extraction rules. While the mathematical complexity is moderate, the empirical depth and systematic analysis of the benchmark’s pipeline are very strong. The findings have important implications for the reliability of published LLM evaluation scores in finance.

Research Flowchart

  flowchart TD
    A[Research Goal: Audit FLARE Scoring Pipeline] --> B{Key Methodology: Score Invariance Analysis};
    B --> C[Data/Inputs: FLARE Released Headlines Scores, 204-Cell Factorial, 46-Cell Diagnostic, Financial PhraseBank Generations];
    C --> D[Computational Processes: Compare FLARE vs. lm-evaluation-harness Scoring Rules, Calculate Ambiguity Rate, Coverage, Floor (floorbench)];
    D --> E{Key Finding 1: Scores Property of Answer Extraction Rule, Not Model};
    E --> F[Key Finding 2: Scores Differ by up to 0.4216 Accuracy Based on Rule];
    F --> G[Key Finding 3: Many Models Below Majority Floor under One Rule, Clear under Another];
    G --> H[Outcome: FLARE Scores Need Ambiguity Rate, Coverage, Floor for Proper Interpretation];

Authors: Jaehyun Kim, Hoi Ying Wong

Abstract

Shortfall aversion describes a form of consumption habit formation in which an individual’s utility depends on consumption relative to the historical consumption peak. This study develops a duality framework for a broad class of utility functions, with and without discounting, avoiding the need for determining the local optimality required by the HJB approaches commonly adopted in the literature. We show that shortfall aversion can instead be incorporated directly into the dual problem through a singular-control mechanism characterized by a variational inequality. Furthermore, we establish that the resulting value function strictly dominates those obtained in the existing literature while remaining attainable under an admissible optimal control. We demonstrate that the suboptimality of existing solutions stems from their myopic treatment of updates to the consumption peak, whereas our duality approach enlarges the admissible class to treat the consumption peak as a separate nondecreasing control and determines its optimal evolution through duality, accounting for future consumption opportunities. These results reveal a previously unrecognized advantage of the dual formulation and provide a more general and flexible framework for consumption optimization under shortfall aversion.

Complexity vs Empirical Score

• Math Complexity: 9.0/10
• Empirical Rigor: 2.0/10
• Quadrant: Lab Rats — theoretically deep, empirically untested

Why this score: This paper presents a highly mathematical and theoretical framework for consumption optimization under shortfall aversion, introducing a novel duality approach. While the mathematical rigor is strong, there is no empirical validation or backtesting presented in the summary or excerpt, placing it firmly in the ‘Lab Rats’ quadrant. The novelty of the duality framework and its ability to address limitations of existing HJB approaches are significant.

Research Flowchart

  flowchart TD
    A[Research Goal: Optimize Consumption under Shortfall Aversion] --> B{Methodology: Duality Framework Development};
    B --> C[Key Mechanism: Singular Control via Variational Inequality];
    C --> D{Computational Process: Optimal Control for Consumption & Peak Evolution};
    D --> E[Outcome 1: Value function strictly dominates existing solutions];
    E --> F[Outcome 2: Optimal treatment of consumption peak (nondecreasing control)];
    F --> G[Outcome 3: More general & flexible framework for consumption optimization].

Authors: Hrishikesh Singh, Aarti Sharma, Millie Pant

Abstract

In modern financial markets, more than 3/4th of the large volume trade is executed by algorithmic trading based on advanced trading strategies to attain higher precision and speed than traditional human traders. The conventional strategies focused on price-based indicators are inadequate in capturing the market dynamism arising from real-world events such as social sentiments, political instability, or sudden economic shifts. With the rise of Deep Reinforcement Learning (DRL) and data availability, there is a requirement for more efficient algorithms that can incorporate these external factors in trading strategies. DRL has shown potential in simulating real-world scenarios and in simplifying complex decision-making since DRL agents can automate the trading process i.e. executing trades in a manner comparable to human traders. They have shown reliable performance since they can update their policies based on risk and quantitative factors. In this paper, we propose a DRL approach based on partial and deep-infusion of sentimental analysis. It features distinct policy designs of actor-critic-based agents: only policy optimization (A2C & PPO) and policy optimization along with Q-learning principles (DDPG, TD3 & SAC). Our findings reveal, that the agents based on policy optimization combined with Q-learning, significantly and consistently demonstrate improved performance upon sentiment deep-infusion in profit maximization as well as risk-averse trading.

Complexity vs Empirical Score

• Math Complexity: 6.5/10
• Empirical Rigor: 6.0/10 (capped from 7.0: transaction costs not modelled; see penalty flags)
• Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper presents a solid DRL approach with a clear methodology for integrating sentiment analysis. It demonstrates good empirical rigor through its data selection and comparative analysis of DRL agents, while also incorporating advanced mathematical concepts from reinforcement learning.

Research Flowchart

  flowchart TD
    A[Research Goal: Develop DRL for Stock Trading incorporating External Factors] --> B{Key Methodology: DRL with Partial/Deep Sentiment Infusion};
    B --> C[Agent Designs: Actor-Critic (A2C, PPO) & Actor-Critic + Q-Learning (DDPG, TD3, SAC)];
    C --> D[Data/Inputs: Financial Market Data + Sentiment Analysis Data];
    D --> E[Computational Processes: DRL Training & Policy Optimization];
    E --> F[Key Findings: Policy Optimization + Q-Learning Agents significantly improve performance with sentiment deep-infusion];
    F --> G[Outcomes: Improved Profit Maximization & Risk-Averse Trading];

Authors: Aleksandar Arandjelovic, Patrick Cheridito

Abstract

Stratification can reduce the estimator variance in Monte Carlo simulations by partitioning the sample space according to a low-dimensional factor and sampling separately within the resulting strata, making it well suited to the valuation of path-dependent claims in high-dimensional financial models. Its performance, however, depends critically on the choice of the stratification factor, and effective low-dimensional factors are generally difficult to identify in complex continuous-time models. Neural networks provide a flexible approximation class for this purpose and have proved useful in related variance-reduction problems in quantitative finance. In this paper, we study the construction of neural stratification factors for derivative pricing under proportional allocation in continuous-time models. Since the numerical construction of a stratification factor typically involves approximation, we establish continuity results for the associated conditional expectations and, more generally, conditional distributions, and provide several counterexamples showing the limits of this continuity theory. Combined with density results for neural networks, our findings yield a universal approximation theorem for admissible stratification factors. We conclude with numerical experiments involving a constant-maturity-swap range accrual in a two-factor Gaussian short-rate model and a worst-of autocallable in a Wishart stochastic covariance model, where the learned factors outperform all considered benchmarks.

Complexity vs Empirical Score

• Math Complexity: 8.5/10
• Empirical Rigor: 7.0/10
• Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: This paper presents a highly mathematical treatment of neural stratification factors, including continuity results and a universal approximation theorem. It is complemented by solid numerical experiments on complex derivatives, demonstrating strong empirical performance. The blend of theoretical depth and practical application is commendable.

Research Flowchart

  flowchart TD
    A[Research Goal: Optimize Derivative Pricing with Stratification] --> B{Problem: Identifying Effective Stratification Factors in Complex Models};
    B --> C[Methodology: Develop Neural Network-based Stratification Factors];
    C --> D{Computational Process: Neural Network Training for Factor Identification};
    D --> E[Data/Inputs: Continuous-Time Financial Models (e.g., Gaussian Short-Rate, Wishart Stochastic Covariance)];
    E --> F[Key Findings: Universal Approximation Theorem for Admissible Stratification Factors];
    F --> G[Outcomes: Learned Factors Outperform Benchmarks for Derivative Pricing (e.g., CMS Range Accrual, Worst-of Autocallable)];

Authors: An Chen, Christian Laudagé, Shihao Zhu

Abstract

We study optimal collective investment subject to a Value-at-Risk (VaR)-type constraint that accounts for uncertainty about future regulatory requirements. The constraint incorporates several possible wealth thresholds and imposes a minimum weighted average probability of meeting them. We characterize the optimal terminal wealth, which combines regions of constant wealth at regulatory thresholds with Merton-type investment regions. Unlike a classical VaR constraint with only one threshold, our formulation generates protection against several thresholds. For two possible thresholds, we identify when both are protected and when protection concentrates on the larger one. Numerical results show that the regulatory constraint can substantially reshape the optimal payoff and generate large welfare losses relative to unconstrained investment when the required compliance probability is high, even the smaller wealth threshold is high, or initial capital is limited. In heterogeneous investor pools, welfare losses depend strongly on the composition of the risk aversions and the Pareto weights governing risk sharing. Losses are particularly pronounced when highly risk-averse investors dominate the pool. Moreover, investors with smaller capital contributions can bear disproportionately large shares of the total welfare loss, while larger contributors may bear less than their capital-proportional share. Thus, regulatory uncertainty not only reduces aggregate welfare but can also create substantial disparities in how its costs are distributed among investors.

Complexity vs Empirical Score

• Math Complexity: 8.5/10
• Empirical Rigor: 6.0/10
• Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper presents a highly mathematical framework using martingale methods for optimal portfolio choice under a novel VaR-type constraint. While the empirical section is numerical rather than data-driven, it is well-executed and provides strong insights. The core contribution lies in extending VaR constraints to multiple regulatory thresholds and analyzing welfare distribution in heterogeneous investor pools.

Research Flowchart

  flowchart TD
    A[Research Goal: Optimal Collective Investment under Regulatory Uncertainty] --> B{Methodology: Stochastic Control & VaR-type Constraint};
    B --> C[Inputs: Regulatory Uncertainty (Multiple Wealth Thresholds), VaR Compliance Probability, Investor Risk Aversions, Initial Capital];
    C --> D{Computational Process: Characterize Optimal Terminal Wealth & Analyze Welfare Impacts};
    D --> E[Key Finding 1: Optimal Terminal Wealth = Constant Wealth at Regulatory Thresholds + Merton-type Regions];
    E --> F[Key Finding 2: Regulatory Constraint Reshapes Payoff, Causes Substantial Welfare Losses (High Compliance Prob, High Thresholds, Limited Capital)];
    F --> G[Key Finding 3: Welfare Losses & Distribution Depend on Investor Heterogeneity (Risk Aversion, Pareto Weights, Capital Contribution)];