About this issue
The first issue of Temerity Holdings’ Quant Trading & Market Microstructure brief — a primer that orients the systematic-trading research space across four pillars (microstructure, signal evaluation, backtesting, and machine learning) plus a roundup of recent academic work. Section syntheses are AI-generated; every source is linked for primary reading.
Quant Trading and Market Microstructure — Vol. 1, Issue 1
Introduction
Systematic trading lives at the intersection of four disciplines that rarely get studied together: the microstructure theory of how orders become prices, the signal research that proposes what to trade, the statistical methodology that decides whether a signal is real, and the machine learning toolkit that increasingly does the searching. A quant who is strong in one and weak in the others tends to produce backtests that look spectacular and trade terribly. This primer is built to orient all four at once.
The throughline of the issue is adverse selection in two forms. In the market, it is the toxic order flow that drives liquidity providers to widen or withdraw — the microstructure problem. In research, it is selection bias: the near-certainty that, having tested enough signals, you will “discover” one that is pure noise. The most useful papers below — from Easley–López de Prado–O’Hara on flow toxicity to Harvey–Liu–Zhu on the factor zoo to Bailey–López de Prado on the deflated Sharpe ratio — are, at bottom, about not being the sucker in either game.
After reading this issue you should have a working map of the field: the vocabulary of spreads and order-flow imbalance, the statistical hurdles a signal must clear before it deserves capital, the specific ways backtests lie, and a grounded sense of where machine learning genuinely adds value versus where it mostly adds overfitting. Each section synthesizes across its sources and then points to the two or three worth reading in full.
In This Issue
| Section | Sources |
|---|---|
| Market Microstructure Fundamentals | 7 |
| Signal Design & Evaluation | 4 |
| Backtesting Methodology | 4 |
| Machine Learning in Trading | 5 |
| Academic Research Roundup | 4 |
Market Microstructure Fundamentals
Synthesis by Claude Opus 4.7 — May 2026
Market microstructure is the study of the purest form of financial intermediation: the trading of an asset that is not transformed, merely transferred from one investor to another. Hans Stoll’s survey frames the whole field around a single object — the bid-ask spread — and its decomposition into three sources of cost. Order-processing costs are the mechanical floor (quoting, routing, clearing). Inventory costs compensate liquidity suppliers for bearing risk on unbalanced positions. And adverse-selection costs are the dealer’s protection against better-informed counterparties: as Bagehot argued in 1971, the losses dealers take to informed traders must be offset by profits from uninformed traders, or the market shuts down. The Glosten–Milgrom and Kyle (1985) models formalize this — every trade nudges the dealer’s belief about value, so buys ratchet quotes up and sells ratchet them down. Stoll’s practical contribution is the realized-versus-quoted spread decomposition: if quotes move by fraction δ in response to a trade, the share of the spread that is adverse selection can be backed out from how much of the quoted spread liquidity suppliers actually earn (on 1,706 NYSE stocks in 1998, the quoted half-spread averaged 7.87 cents but the traded half-spread only 3.74 cents — the gap is the information component).
Maureen O’Hara’s “High Frequency Market Microstructure” argues that this classical apparatus is breaking down at machine speed, and that microstructure matters more, not less, as a result. Three shifts stand out. First, speed is now informational — being informed means seeing and acting on prices before competitors, so “to be uninformed is to be slow.” Second, trade direction no longer signals information because execution algorithms shred a parent order into many small child orders that mostly execute passively: in an ITG study of 243,772 parent orders, each generated 55.3 child trades on average, and the aggregate fill mix was 65.5% passive and only 12.6% aggressive. A parent buy therefore shows up in the tape mostly as small sells. Third, the classical estimators fail — realized spreads measured five minutes after a trade are often negative (“five minutes is a lifetime”), and PIN estimation breaks numerically. O’Hara’s prescription is to abandon the clock for the volume clock and to measure toxicity directly.
That measure is VPIN — the Volume-Synchronized Probability of Informed Trading (Easley, López de Prado, O’Hara). Group sequential volume into equal-sized buckets, classify each bucket’s volume into buy- and sell-initiated portions, and track the order imbalance smoothed over roughly 50 buckets. It requires no unobservable parameters and updates in volume time. Its signature result: toxicity spiked ahead of the May 2010 Flash Crash, consistent with the mechanism O’Hara describes — one-sided flow forces market makers to hit position limits and withdraw, producing episodic illiquidity. Jonathan Kinlay’s practitioner write-up shows the idea operationalized on E-mini futures, where a regime-switching model on signed VPIN lifts explanatory power to an R² of 27% (versus ~7% for a linear ARMA) and the prior bucket’s signed flow correlates around 0.5 with the next midprice change.
The complementary lens is price impact. Cont, Kukanov and Stoikov show that over short horizons, midprice changes in equities are driven by order-flow imbalance (OFI) — the net change in supply and demand at the best bid and ask — not by trade volume. The relation is strikingly linear (average R² of 65%, versus 32% for trade imbalance) with a slope inversely proportional to market depth; because depth at the open is roughly half its daily average, price impact is about twice as large at the open. They also show the famous “square-root law” of impact-versus-volume is largely a statistical artifact of aggregation, noisier and less robust than the OFI relation. Kyle and Obizhaeva’s microstructure-invariance work supplies the cross-asset scaling skeleton: treat trading as a sequence of “bets” arriving in business time, and quantities that vary wildly across assets become near-constants per unit of business time. Their scaling laws (bet arrival ∝ trading-activity^2/3; bet size and percentage cost ∝ activity^±1/3) let you normalize spreads, sizes and impact across equities, futures and bonds from a single benchmark.
That OFI lens is not just an equities curiosity. A November 2025 Federal Reserve note shows the same dynamic in U.S. Treasuries: during the April 2025 tariff turbulence, a surge to over $2 billion of net sell-side imbalance produced a large yield move, while a comparable volume spike two days later — but with imbalance staying bidirectional and under $1 billion — barely moved prices. The lesson generalizes across the issue: directional, persistent order flow hitting thin depth is what amplifies prices; raw volume is secondary. For a systematic desk, OFI and VPIN are the implementable signal primitives, and Kyle–Obizhaeva scaling is the discipline that keeps position sizing honest across instruments.
Sources & Further Reading
| Title | Publication | ~Read | Why It’s Worth Reading |
|---|---|---|---|
| High Frequency Market Microstructure | Journal of Financial Economics (2015) | ~35 min | O’Hara’s field-defining account of why HFT broke the classical models — and the data (cancel rates, child-order fill mixes) behind it |
| Flow Toxicity and Liquidity in a High-Frequency World (VPIN) | Review of Financial Studies (2012) | ~45 min | The source of the volume clock and VPIN; the canonical toxicity measure feeding TH’s indicator work |
| The Price Impact of Order Book Events | Journal of Financial Econometrics (2014) | ~35 min | Clean, implementable result: OFI predicts short-horizon price moves linearly, slope ∝ 1/depth, R²≈65% |
| Market Microstructure Invariance: Empirical Hypotheses | Working paper (Kyle & Obizhaeva) | ~40 min | The cross-asset scaling laws for bet size, arrival rate and trading cost — the backbone of disciplined sizing |
| Market Microstructure (survey) | Handbook of the Economics of Finance (2003) | ~30 min | The clearest exposition of spread decomposition (order-processing / inventory / adverse selection) and the Roll estimator |
| Measuring Toxic Flow for Trading & Risk Management | Quantitative Research & Trading (Kinlay) | ~18 min | Practitioner implementation of VPIN with regime-switching and a worked futures example, with code |
| Order Flow Imbalances and Amplification of Price Movements: U.S. Treasury Markets | Federal Reserve FEDS Note (Nov 2025) | ~12 min | Fresh, free evidence that directional imbalance (not volume) drives amplification — the April 2025 Treasury episode |
Go deeper: O’Hara, High Frequency Market Microstructure for the conceptual reframing of the whole field, and Cont–Kukanov–Stoikov for the single most directly tradeable microstructure result.
Signal Design & Evaluation
Synthesis by Claude Opus 4.7 — May 2026
A trading signal is a hypothesis about future returns, and the central discipline of signal evaluation is refusing to believe your own hypotheses too easily. Harvey, Liu and Zhu’s “…and the Cross-Section of Expected Returns” is the foundational text here. They count 316 factors published in the literature — Cochrane’s “factor zoo” — and point out that with that much multiple testing against the same data, the usual significance cutoff of t > 2.0 is meaningless: many “discovered” factors are false positives that cleared a bar set for a single test. Applying multiple-testing corrections (Bonferroni, Holm, and the false-discovery-rate-based Benjamini–Hochberg–Yekutieli), they estimate that between 80 and 158 of 296 published significant factors would be reclassified as false discoveries. Their headline recommendation is that a newly discovered factor today needs a t-statistic above 3.0 (p ≈ 0.27%) — and because the count of tests keeps rising, so does the hurdle, reaching roughly 3.8 by 2012 and a projected 4.0 by 2032. Even a theory-derived factor, they argue, no longer deserves the old t > 2.0.
Clearing the hurdle is necessary but not sufficient, because signals decay once they are known. The literature’s stylized fact (McLean & Pontiff) is that anomaly returns fall by roughly half after publication — investors learn and arbitrage them away. A provisional 2025 working paper, “Not All Factors Crowd Equally,” pushes on the mechanism: it models alpha as decaying hyperbolically under a crowding equilibrium and argues that mechanical factors easily replicated from public data (momentum, reversal) decay predictably, while judgment-based factors (value, quality) resist crowding because signal ambiguity creates an information barrier. Its more provocative claim is that crowding predicts crashes, not means — making it a tail-risk input rather than an alpha source. (Treat its specific figures cautiously: the preprint was withdrawn for major revision shortly after posting; the conceptual framing is the takeaway, not the numbers.)
Before you can even claim a signal produces alpha, you have to net out the factor exposures you are already being paid for. AQR’s “Measuring Portfolio Factor Exposures” is the practitioner’s guide: regress portfolio returns on the standard factors (market, value, momentum, size), and what naïvely looks like alpha usually resolves into ordinary factor beta. In their worked example, roughly 4% of a long-only value-and-momentum portfolio’s return decomposes cleanly into HML, UMD and SMB contributions — return that a single-factor CAPM lens would have miscredited as skill. The piece is most valuable for its catalogue of interpretation traps: judge betas by their t-statistics, not their magnitude; don’t compare betas across portfolios of different volatility without normalizing per unit of risk; always check the regression’s R²; and beware that academic factors are gross of costs and span a far broader (small-cap-heavy) universe than any investable portfolio, so a strategy can capture a factor perfectly and still post negative regression alpha purely from implementation frictions. AQR’s framing line is the one to remember: “without the proper model, rewards for factor exposures may be misconstrued as alpha.”
The constructive counterpoint is meta-labeling. Once you have a vetted primary signal that sets direction, López de Prado’s two-stage design adds a secondary model that decides whether and how much to act on each signal. Hudson & Thames’ study pairs it with the triple-barrier labeling method (label each trade by which of three barriers — profit-take, stop-loss, or time horizon — is hit first) and reports that meta-labeling roughly tripled a mean-reversion strategy’s out-of-sample Sharpe (0.45 → 1.48) and improved a trend-follower’s, chiefly by filtering low-conviction signals and cutting false positives. A useful incidental finding: sampling on dollar bars rather than time bars produced returns far closer to normal (an order-of-magnitude better Jarque–Bera statistic), which matters for everything downstream. Together these four sources define a workflow for TH: hold signals to a high statistical bar, decompose apparent alpha against known factors, expect post-discovery decay, and use meta-labeling to size and filter what survives.
Sources & Further Reading
| Title | Publication | ~Read | Why It’s Worth Reading |
|---|---|---|---|
| …and the Cross-Section of Expected Returns | Review of Financial Studies (2016) | ~40 min | The multiple-testing case for a t > 3.0 hurdle, with the false-discovery counts that make it concrete — essential reading |
| Not All Factors Crowd Equally: Modeling, Measuring, and Trading on Alpha Decay | arXiv working paper (2025, withdrawn for revision) | ~35 min | A recent model of how and why factors decay and crowd unevenly; read for the framing, not the (provisional) numbers |
| Does Meta-Labeling Add to Signal Efficacy? | Hudson & Thames | ~15 min | A concrete, reproducible test of triple-barrier labeling + meta-labeling, including the dollar-bars-vs-time-bars result |
| Measuring Portfolio Factor Exposures: A Practical Guide | AQR | ~25 min | The practitioner’s checklist for separating real alpha from factor beta, and the interpretation traps that defeat most people |
Go deeper: Harvey–Liu–Zhu for the discipline of not believing a signal too easily, and AQR’s factor-exposure guide for the practical mechanics of attributing returns.
Backtesting Methodology
Synthesis by Claude Opus 4.7 — May 2026
The single most important number missing from almost every backtest is the count of strategy variants the researcher tried before reporting the winner. That is the argument running through Bailey and López de Prado’s work, and its blunt corollary is that “a backtest where the researcher has not controlled for the extent of the search involved is worthless, regardless of how excellent the reported performance might be.” The reason is selection bias: search enough configurations and the best in-sample Sharpe will look impressive even if every candidate is pure noise. A companion demonstration (“Statistical Overfitting and Backtest Performance,” using their online SEBO tool) makes the point viscerally — running a trivial strategy with a handful of parameters on purely random price data, exhaustively searching ~55,000 parameter combinations, produces in-sample Sharpe ratios centered around 0.9 while the out-of-sample ratios sit exactly where they should, at zero. There is, the authors stress, no Sharpe threshold or haircut that can be considered safe; adding even one more parameter pushes the overfit Sharpe to any desired level. Worse, because financial series have memory, overfitting doesn’t merely revert to zero out-of-sample — the extreme in-sample patterns get “undone,” so overfit strategies tend to lose money, not just fail to make it.
The Deflated Sharpe Ratio (DSR) is the parametric fix. It is a probability — that the true Sharpe exceeds zero — but computed against a threshold that accounts for everything a naïve Sharpe ignores: the number of independent trials N, the variance of the Sharpe ratios across those trials, the sample length, and the non-normality (skew and fat tails) of returns. The key insight, drawn from extreme-value theory, is that the expected maximum Sharpe across N trials grows with N even under zero skill, so that expected maximum becomes the bar the observed Sharpe must clear. Their worked example is instructive: a strategy reporting an annualized Sharpe of 2.5 over five years of daily data is rejected at 95% confidence once the search behind it is disclosed — yet the very same Sharpe would have passed had it emerged from only ~46 independent trials rather than the larger number actually run. Both inflation sources — multiple testing and non-normality — have to be deflated jointly.
The non-parametric, resampling-based family attacks the same problem differently. The Probability of Backtest Overfitting (PBO), via combinatorially-symmetric cross-validation, estimates how often the configuration that is best in-sample lands below the median out-of-sample — a direct, distribution-free measure of overfitting. And purged, embargoed cross-validation addresses why ordinary k-fold cross-validation is invalid on financial data in the first place: observations aren’t independent, features have overlapping lookback windows, and labels span future price paths, so naïve folds leak the future into the past. QuantInsti’s walkthrough explains the two repairs — purge training observations whose label horizons overlap the test set, then embargo a buffer immediately after the test fold sized to the longest feature lookback — and then Combinatorial Purged Cross-Validation (CPCV), which recombines folds to generate many distinct backtest paths instead of one. The payoff is a distribution of outcomes: their SPY example shows a single full-history path with a seductive Sharpe of 3.49 that nonetheless hid a 98% drawdown, where CPCV instead yields a spread of Sharpe ratios that exposes the fragility.
Robot Wealth translates all of this into a practitioner’s idiom, and the translation is mostly about humility. It names the same three biases — look-ahead (using information not yet available, e.g. trading on the day’s close intraday), over-optimization (fitting noise with too many parameters), and data-mining (cherry-picking the market or variant that happened to work) — and prescribes the antidotes: prefer strategies you can explain in one sentence, optimize for broad robust parameter plateaus rather than a single peak, keep in-sample data strictly for discovery (which parameters matter, sensible ranges) and never for measuring performance, and treat any backtest showing “thousands of percent” returns as a red flag rather than a discovery. The convergent message across all four sources: track N, deflate the Sharpe accordingly, validate with purged/combinatorial CV, and let simplicity and out-of-sample honesty do the heavy lifting.
Sources & Further Reading
| Title | Publication | ~Read | Why It’s Worth Reading |
|---|---|---|---|
| The Deflated Sharpe Ratio | Journal of Portfolio Management (2014) | ~30 min | The parametric tool for discounting a Sharpe by the number of trials and non-normality — with a worked rejection example |
| The Probability of Backtest Overfitting | SSRN / Journal of Computational Finance (2015) | ~35 min | Introduces PBO and combinatorially-symmetric CV — a distribution-free measure of how badly a strategy is overfit |
| Cross-Validation in Finance: Purging, Embargoing, Combinatorial | QuantInsti | ~18 min | The clearest explainer of why k-fold CV leaks in finance and how purging/embargoing/CPCV fix it, with a worked example |
| Backtesting Bias: Feels Good, Until You Blow Up | Robot Wealth | ~13 min | An accessible practitioner tour of look-ahead, overfitting and data-mining bias, and the simplicity-first antidotes |
Go deeper: The Deflated Sharpe Ratio for the one correction every backtest should carry, and QuantInsti’s CV guide for the validation machinery that prevents leakage in the first place.
Machine Learning in Trading
Synthesis by Claude Opus 4.7 — May 2026
The benchmark study for machine learning in asset pricing is Gu, Kelly and Xiu’s “Empirical Asset Pricing via Machine Learning,” a head-to-head bake-off of thirteen models on roughly 30,000 U.S. stocks over 1957–2016, using 94 firm characteristics interacted with macro variables for more than 900 predictors. The results are stark and have shaped the field’s expectations. Ordinary least squares with the full predictor set is worse than useless — a monthly out-of-sample R² of −3.46% — while regularized linear models claw back to roughly zero and the nonlinear methods (random forests, gradient-boosted trees, and especially neural networks) reach monthly stock-level R² between 0.33% and 0.40%. Those small R² numbers translate into large economic gains: a value-weighted long-short decile spread built on the best neural net earns an annualized Sharpe of 1.35 versus 0.61 for the regression benchmark — “in some cases doubling the performance.” Three findings recur and matter for any ML effort. The gains come specifically from interaction effects, not mere nonlinearity (a generalized linear model with splines but no interactions manages only 0.19%). Shallow beats deep — performance peaks at three hidden layers and degrades thereafter, a direct consequence of finance’s low signal-to-noise ratio. And the dominant predictors are consistent across every method: price trends (momentum, reversal), liquidity, and volatility. The authors are careful about what this is not: ML delivers better measurements of expected returns, not explanations of economic mechanism.
A 2024 survey from NJIT (“From Factor Models to Deep Learning”) widens the lens, arguing ML reshapes the entire empirical-asset-pricing pipeline rather than just the prediction step: dimension reduction and latent-factor models (autoencoders, FactorVAE), missing-data imputation, graph neural networks that model inter-asset relationships from supply chains and correlation structure, alternative-data fusion (charts-as-images, earnings-call text and audio), and reinforcement learning that outputs portfolio weights directly. But it is equally clear-eyed about the obstacles, which the CFA Institute’s practitioner chapter (by Robeco quants) sharpens into four challenges that make finance uniquely hostile to ML. Signal-to-noise is extremely low — no single feature strongly determines returns. Data is small: even a century of monthly U.S. data is barely 1,200 observations per security, so simpler models on more data usually win. Markets are non-stationary and adaptive — published anomalies decay precisely because the market learns. And human irrationality injects structure that has no analogue in physics or protein-folding. The practical upshot is a discipline of restraint: cross-validation, regularization, early stopping, SHAP-value “smell tests” against domain knowledge, scrupulous point-in-time data to prevent leakage, and a sober accounting of the implementation gap (ML strategies often concentrate their alpha in small-caps and carry high turnover). The CFA chapter’s memorable line: “ensembling is one of the closest things to a free lunch in ML, much like diversification for investing.”
Two sources supply the necessary reality check at the microstructure end, where the gap between statistical accuracy and tradeable signal is widest. Briola, Bartolucci and Aste’s deep limit-order-book study finds that a stock’s tick size governs how predictable it is — large-tick stocks carry a stronger, more robust mid-price signal — but, crucially, that high classification accuracy does not imply a tradeable strategy. They show forecasts with strong F1 and MCC scores that yield zero correctly executed transactions, because the timing of the correct predictions is wrong; conventional ML metrics simply fail to assess forecast quality in the LOB context, and even where the signal is real, exploiting it requires low-latency hardware. Kinlay’s reinforcement-learning execution piece makes the analogous point about optimal execution: a PPO agent conditioning on live order-book state beats Almgren–Chriss and TWAP, but only modestly (about 0.7 basis points versus Almgren–Chriss), with the edge concentrated in large parent orders and adverse, volatile, imbalanced regimes — below roughly 5% of interval volume, “none of this matters” and a closed-form schedule is fine. His recommended deployment is a size-threshold hybrid, not “RL replaces everything.” Read together, the message for TH is balanced: ML genuinely and reproducibly improves cross-sectional return forecasting (the GKX result is robust), but the value is conditional — shallow models, interaction-rich features, relentless overfitting control, and a hard-nosed distinction between forecast accuracy and executable edge.
Sources & Further Reading
| Title | Publication | ~Read | Why It’s Worth Reading |
|---|---|---|---|
| Empirical Asset Pricing via Machine Learning | NBER / Review of Financial Studies (2020) | ~60 min | The benchmark ML-in-finance study; the source of “shallow beats deep,” “interactions are the edge,” and the doubled Sharpe |
| From Factor Models to Deep Learning | arXiv survey (2024) | ~45 min | A map of the full ML asset-pricing frontier (GNNs, autoencoders, RL, alt-data) plus an honest list of obstacles |
| Deep Limit Order Book Forecasting | arXiv (2024) | ~40 min | Shows tick size governs LOB predictability and that high ML accuracy ≠ tradeable signal; ships the open-source LOBFrame |
| AI and Big Data in Investments (Part I) | CFA Institute Research Foundation | ~20 min | A practitioner’s account of why finance is hostile to ML and the specific pitfalls and mitigations that follow |
| Reinforcement Learning for Optimal Execution | Quantitative Research & Trading (Kinlay) | ~20 min | A grounded RL-execution study: where RL beats Almgren–Chriss/TWAP, by how little, and when not to bother |
Go deeper: Gu–Kelly–Xiu for the result that anchors the field, and the CFA practitioner chapter for the discipline that keeps ML from overfitting itself into oblivion.
Academic Research Roundup
Synthesis by Claude Opus 4.7 — May 2026
The recent literature (late 2025 into 2026) is converging on a healthier set of obsessions: honest validation over inflated claims, better inputs over deeper networks, and metrics that measure tradeability rather than mere accuracy. The strongest statement of the first theme is Deep, Deep and Lamptey’s “Interpretable Hypothesis-Driven Trading,” which opens by noting that institutional investors report over 90% of academic strategies fail with real capital even as papers routinely claim 15–30% annual returns. Their contribution is deliberately not a profitable strategy but a validation template: human-readable microstructure hypotheses, a reinforcement-learning agent that learns which to act on, and strict walk-forward testing across 34 independent out-of-sample folds spanning 2015–2024. The honestly reported result is a 0.55% annualized return, a Sharpe of 0.33, and a p-value of 0.34 — not statistically significant, with the authors noting their 34 folds afford only 12% statistical power. The substantive finding survives anyway: the daily-data microstructure signals work only in high-volatility regimes (2020–2024) and vanish in calmer periods, because they need elevated information arrival to function at all. It is a model of what disciplined, non-cherry-picked research looks like.
The architecture debate is maturing toward pragmatism rather than a single winner. Kinlay’s survey on state-space models asks whether Mamba can replace Transformers in high-frequency finance and answers, sensibly, that this is the wrong question. State-space models bring genuine efficiency advantages — O(n) versus O(n²) scaling, with single-digit-millisecond latency on long context windows — that matter enormously when latency is the whole game, but Transformers retain the edge on explicit order-book structure, and the practical frontier is hybrids (Graph-Mamba, FinMamba) rather than wholesale replacement. He flags the real bottleneck as the absence of head-to-head, regime-stress-tested benchmarks on identical LOB datasets — exactly the rigor the Deep et al. paper models.
A cryptocurrency LOB study (BTC/USDT, “Better Inputs Matter More Than Stacking Another Hidden Layer”) presses the “inputs over architecture” theme directly. With proper denoising — a Savitzky–Golay filter — simple models match or beat deep nets: logistic regression and XGBoost both score ~0.728 against DeepLOB’s 0.719, at a fraction of the cost and with far better interpretability. And order-book depth turns out to matter far more than model depth (40 levels yields 0.715 accuracy versus 0.58 at 5 levels). It is a useful empirical counterweight to the reflex that another hidden layer is the answer — and it converges with the tick-size and accuracy-versus-tradeability findings from Briola, Bartolucci and Aste discussed in the Machine Learning section above: a single accuracy-style number can rise while the strategy becomes less tradeable.
The most technically novel entry is KANFormer, which reframes order execution as a survival-analysis problem: instead of classifying price direction over an arbitrary horizon, predict the full time-to-fill distribution of a limit order, treating canceled or expired orders as right-censored. Combining dilated causal convolutions, a Transformer encoder, and Kolmogorov–Arnold network layers on labeled CAC 40 futures data, it outperforms every baseline (a concordance index of 0.72 versus 0.63–0.67 for the next best). Its most quotable result is an ablation: removing queue position collapses discrimination (integrated AUC from 0.76 to 0.57), far more than removing any other feature — a direct, quantitative confirmation that microstructurally-grounded inputs, not architectural depth, are what carry the signal. Taken together, the 2025–26 frontier is reassuringly aligned with the rest of this primer: it rewards honest out-of-sample testing, careful feature and microstructure grounding, and metrics that reflect whether a model can actually trade.
Sources & Further Reading
| Title | Publication | ~Read | Why It’s Worth Reading |
|---|---|---|---|
| Interpretable Hypothesis-Driven Trading: A Walk-Forward Validation Framework | arXiv (Dec 2025) | ~35 min | A template for honest validation — walk-forward, interpretable, and willing to report a statistically insignificant result |
| State-Space Models for Market Microstructure: Can Mamba Replace Transformers? | Quantitative Research & Trading (Kinlay) | ~22 min | A current, balanced read on SSM-vs-Transformer trade-offs and why hybrids, not replacement, are the frontier |
| Exploring Microstructural Dynamics in Crypto Limit Order Books | arXiv (Jun 2025) | ~30 min | Empirical evidence that preprocessing and order-book depth beat architectural complexity for LOB prediction |
| KANFormer: Predicting Fill Probabilities via Survival Analysis in LOBs | arXiv (Dec 2025) | ~30 min | A novel survival-analysis framing of execution, with an ablation proving queue position is the decisive input |
Go deeper: Interpretable Hypothesis-Driven Trading as a model of research honesty, and KANFormer for a genuinely novel reframing of execution as a survival-analysis problem.
My Notes
Added after reading