From papers to positions: what four AI models found reading 468 academic studies

Overview

I gave four frontier language models from four different developers (Opus 4.8, Codex 5.6 Sol, Grok 4.5, and GLM 5.2) the same 468 academic trading-strategy papers, distilled to their themes and rule-mapped onto liquid futures, and asked each the same question: which of these effects survive as futures algorithms, and how should they be combined? The four reviews were mutually blind. I checked the outputs for fabricated references, and had the numbers that carry the most weight fact-checked against the source papers.

The result is a set of 10 futures themes on which the models independently converged: a layer of agreement between models sitting on top of the agreement between papers each model was asked to find. Below are those themes, the ideas each model saw alone, the handful of real disagreements, and a ranked shortlist of six candidate algorithms, with the crowding and in-sample caveats attached.

How this was done

Most academic anomalies never become futures algorithms because the literature is written on cash equities or on idealised long/short portfolios. It says nothing about roll, margin, or capacity, which between them decide whether an effect is tradable on a handful of liquid contracts. This experiment treats that translation problem directly, with four independent language models as the readers.

The corpus is 468 structured papers covering equities, multi-asset, commodities, rates, FX, options, and crypto. Every reported Sharpe, return, or drawdown lives only in free text, so each is an unverified in-sample claim by the paper that reported it. A deterministic pass distilled each paper to a compact record and rule-mapped it onto the most natural liquid futures class: equity-index (ES/NQ/RTY/YM), rates (ZN/ZB/ZF/GE), commodity (CL/GC/NG/HG/ZC), FX (6E/6J/6B/DX), volatility (VX), or crypto (BTC/ETH). The mapping was deliberately generous. I wanted every theme considered as a futures candidate, including the ones written on cash equities.

Each model received a byte-identical bundle and prompt, and each returned the same four sections: unique insights, cross-confirmation across papers, novel algorithm designs, and a single best idea. Opus 4.8 ran as an agent; the other three ran through their own command-line tools. None of the four saw the prompt or output of any other. Afterwards, an automated check confirmed that every cited paper identifier resolved to a real record, and a separate adversarial agent fact-checked the 10 numbers that carry the most weight.

Where four models independently agreed

Four models trained on different data, by different organisations, read the same evidence separately and produced almost the same short list. That’s the central result. Table 1 sets it out.

What’s striking is where the emphasis fell. On trend (M1), the panel spent little time on the signal itself, which is the corpus’s most-cited and most-crowded effect ([1]), and a lot on the machinery around it: closed-loop volatility targeting rather than open-loop inverse-vol sizing, where optimal leverage sits near $\mu/\sigma^2$ ([18]; [19]); a changepoint gate that cuts risk at turning points, where trend systems bleed ([2]); and a refusal to hold the naive full-short equity leg in bear regimes, where the optimal exposure is near zero rather than $-1$ ([3]).

The models also agreed that stacking similar indicators is fake diversification. Moving averages, breakouts, and the rest of the trend-filter family encode nearly the same weighted sum of past returns ([22]), and adjacent lookbacks are redundant ([23]). Real diversification comes from economically different markets and separated timescales.

On volatility (M2), all four re-derived the same structural point. The VIX-futures basis has little power to forecast spot VIX, but it does forecast VIX-futures returns ([4]). What trades is the roll of the contract itself, so a short-VX position is a carry trade, gated by the term structure and sized inversely to implied volatility ([5]). Every model attached the same warning: crowded, and catastrophically left-tailed.

The commodity theme (M3) is the one place where the futures structure itself is the edge. Basis-momentum and the front-of-curve spread ([6]; [7]) can’t be expressed in cash or ETF form, because roll yield and curve shape are the product. The session and flow effects (M4, M5, M8) are similar: overnight returns dominate the trading day across roughly two dozen markets ([9]; [8]), and ES and NQ trade almost around the clock, so a futures book can hold the paid session and stand flat through the unpaid one. Futures capture that effect better than the cash market that produced it.

What each model saw that the others missed

Consensus is half the value of a panel. The rest is what each reader caught on its own.

Opus 4.8 produced the synthesis I liked most, “Convex Carry”: pair the short-VX carry with a fast trend overlay sized so its crisis payoff covers the 95th-percentile loss of the volatility sleeve. Separately, they’re two crowded, dangerous trades. Together, the trend leg is a self-funding tail hedge, and SPAN margin offsets make the book capital-efficient.

Codex 5.6 Sol gave the cleanest decomposition. A futures curve carries two distinct alphas, slow carry continuation and fast front-spread reversal, and they should be traded as separate sleeves rather than blended. Codex called it the “Curve-Speed Barbell”: a weekly first-versus-second calendar-spread reversal bolted onto the monthly carry sleeve ([24]).

Grok 4.5 was the most operationally ruthless of the four. Its hardest check: naive market-making on CME futures nets roughly zero after adverse selection across 13 contract roots ([25]), which kills any earn-the-spread idea before it starts. It also noted that listwise learning-to-rank beats independent per-asset scoring for cross-asset momentum ([26]).

GLM 5.2 had the two micro-insights I hadn’t seen before. Statistical-arbitrage residuals revert about 2.4x faster in rank space than in name space ([27]), which motivates a rank-space index-futures pairs book. And 52-week-high proximity beats 12-1 momentum in small-cap and non-US panels, so the edge lives in RTY and NQ while ES is the beta backdrop ([28]).

Where they disagreed

The disagreements are narrow. The clearest is what to build first. Grok and GLM both chose the regime-gated trend core, with commodity basis-momentum second and VX carry a small third. Opus and Codex preferred a single specialist sleeve. I read that as a difference in risk appetite: the trend core is the higher-capacity, more durable choice, while the specialist sleeves offer more edge per dollar at lower capacity.

On the equity short, all four rejected the naive full-short, but their fixes differed, from a small negative band to a hard zero to a positive deep-loser tilt during recoveries. When four models independently reject the same construction and propose four different replacements, the construction is the problem.

A futures book, assembled from four reviews

No single model laid it out in one place, but the four reviews assemble into one multi-sleeve book. Every sleeve below is backed by multiple papers and by at least three of the four models:

• Core. Regime/changepoint-gated, volatility-targeted cross-asset trend, with long lookbacks, concave exposure, and an equity-short cap in crises.
• Sleeve two. Commodity basis-momentum and front-curve carry, optionally with the fast weekly calendar-spread reversal barbell.
• Sleeve three. Term-structure-gated short-VX carry with a hard kill-switch, ideally in the Convex Carry form where a trend overlay covers its tail.
• Overlays. The overnight/EU-open harvester, gamma-conditioned intraday ES/NQ, turn-of-month flow timing, the FOMC strip, and the FX carry-to-PPP switch, each on a small risk budget.
• Risk bus. Closed-loop volatility targeting plus a breadth and correlation crash filter ([29]; [30]).
• Governance. Backtest-overfitting probability, deflated Sharpe ratios, and walk-forward consistency, applied to every sleeve ([21]; [31]).

Six candidate algorithms, ranked

Ranked by paper-level cross-confirmation and model-level agreement, penalised for crowding and fragility.

1. Regime-gated cross-asset trend core. All four models; the highest-capacity, least glamorous winner. The edge is in the risk machinery ([2]; [32]). The CTA space is crowded, so survival rests on diversification and on restraint at changepoints, and costs have to stay low.
2. Convex Carry (Opus). Short-VX carry tail-hedged by a trend overlay, sized jointly, gated hard on contango ([4]; [5]). Short volatility is negatively skewed and crowded, so validation must run through 2018, 2020, and 2022.
3. Commodity Curve-Speed Barbell (Codex). Slow front-curve carry plus a fast weekly front-spread reversal as separate sleeves ([6]; [24]), executed as calendar spreads. Capacity is limited in NG, HG, and ZC.
4. Rank-space index-futures stat-arb (GLM). A mean-reversion book that ranks by normalised price to re-identify the residual, exploiting the faster rank-space reversion ([27]). Index-pair spreads carry factor mismatch and need dynamic re-estimation.
5. Overnight harvester with an FOMC overlay. Hold ES/NQ only in the paid overnight window ([8]), lever the pre-FOMC drift ([12]), and short front-VX for the post-FOMC crush ([13]). Gap risk and timestamp sensitivity are real.
6. Gamma-conditioned intraday ES/NQ. The highest edge per trade of the six, and by some distance the most fragile: capacity is low ([10]; [11]), the growth of zero-days-to-expiry options has changed dealer positioning since the sample periods, and a published replication already finds the effect weaker ([33]).

The panel was unanimous on what not to build first: unfiltered short volatility, naive futures market-making ([25]), an equity factor-zoo mapped onto ES ([20]), high-frequency crypto without a punishing cost model ([34]), and calendar seasonality as a core rather than an overlay.

Fact-checking the headline numbers

An independent agent checked the 10 numbers that carry the most weight against the source papers. Nine of the 10 traced, word for word, to figures the papers state themselves: the turn-of-month concentration ([16]), the overnight window’s return and Sharpe ([8]), the VIX-basis result ([4]), the roughly 18% basis-momentum spread ([6]), the 2.4x rank-space ratio ([27]), the near-zero net from market-making ([25]), the 36.6% crisis return of an equity-capped trend variant ([32]), the near-zero optimal bear-regime exposure ([3]), and the roughly 49-basis-point pre-FOMC drift ([12]). No mis-citations were found.

One claim was overstated. The line that “76-78% of crisis alpha” comes from beta-timing ([35]) mixes two things. Those figures describe the beta-timing sleeve’s share of the strategy’s total risk and return; the crisis property is shown separately, through downside correlations. I’ve corrected that claim to the weaker, accurate form.

So the numbers are faithful summaries of what each paper claims. But every one is a single-paper, in-sample result, often gross of cost, from a short or idiosyncratic window. Treat them as hypotheses to re-test.

The caveats that matter

Three things constrain everything above.

The edge the panel kept pointing at lives in construction. The raw anomalies are known and crowded; what remains is how exposure is sized and gated, which is the part hardest to arbitrage away and easiest to get wrong.

The corpus is roughly 60% equity and cross-sectional, which inflates the apparent case for equity-index momentum. The strongest recommendations are the futures-native effects, where the structure gives a real edge. Single-name factors tend to die when compressed into one index contract.

And the four models share training data and priors, so their agreement is correlated to an unknown degree. A theme absent from all four might be weak, or might simply be underrepresented in what they were trained on. Independent agreement is still evidence, even when the agents are imperfectly independent.

What to do next

Four things follow.

Reproduce before trusting. Rebuild each sleeve on point-in-time, roll-adjusted futures data with realistic costs, and put it through the governance battery the panel itself demanded, treating every cited number as a hypothesis.

Test the whole architecture, since the strongest claim here is that jointly sizing premia that move in opposite directions under stress, and combining decorrelated sleeves under one volatility controller, beats running any sleeve alone. The 2018, 2020, and 2022 stress episodes are the discriminating experiments.

Spend the first research budget on the futures-native effects (basis-momentum, VX carry, the overnight session) before competing with the entire CTA industry on equity-index trend.

And widen the panel. The protocol is cheap, so extending the corpus and re-running it as new models arrive would show whether the shortlist moves, and which way.

References

1. Akindynos-Nikolaos Baltas & Robert Kosowski (2013). Momentum Strategies in Futures Markets and Trend-Following Funds. https://ssrn.com/abstract=1968996

2. Kieran Wood, Stephen Roberts & Stefan Zohren (2021). Slow Momentum with Fast Reversion: A Trading Strategy Using Deep Learning and Changepoint Detection. https://arxiv.org/abs/2105.13727

3. Valeriy Zakamulin (2026). Rethinking Trend Following: Optimal Regime-Dependent Allocation.

4. David P. Simon & Jim Campasano (2013). The VIX Futures Basis: Evidence and Trading Strategies.

5. Divyanshu Verma (2025). Volatility as an Asset Class: Carry, Convexity, Crash Risk, and Hedge Fund Implementation.

6. Martijn Boons & Melissa Prado (2019). Basis-momentum. https://ssrn.com/abstract=2587784

7. Ana-Maria Fuertes, Joëlle Miffre & Georgios Rallis (2010). Tactical allocation in commodity futures markets: Combining momentum and term structure signals. http://ssrn.com/abstract=1127213

8. Oleg Bondarenko & Dmitriy Muravyev (2021). Market Return Around the Clock: A Puzzle. https://ssrn.com/abstract=3596245

9. Michael Cliff, Michael J Cooper & Huseyin Gulen (2008). Return Differences between Trading and Non-trading Hours: Like Night and Day. https://ssrn.com/abstract=1004081

10. Guido Baltussen, Zhi Da, Sten Lammers & Martin Martens (2020). Hedging demand and market intraday momentum. https://ssrn.com/abstract=3760365

11. UBS Investment Bank (2020). Intraday Trend.

12. David O. Lucca & Emanuel Moench (2013). The Pre-FOMC Announcement Drift. https://ssrn.com/abstract=1923197

13. Adrian Fernandez-Perez, Bart Frijns & Alireza Tourani-Rad (2014). When No News is Good News – The decrease in Investor Fear after the FOMC announcement. https://ssrn.com/abstract=2525991

14. Pedro Barroso & Pedro Santa-Clara (2015). Beyond the Carry Trade: Optimal Currency Portfolios. https://ssrn.com/abstract=2041460

15. Fabian Ackermann, Walt Pohl & Karl Schmedders (2016). Optimal and Naive Diversification in Currency Markets. https://ssrn.com/abstract=2184336

16. Otto van Hemert (2014). The MOM-TOM effect: Detecting the market impact of CTA trading. https://ssrn.com/abstract=2515900

17. Daniel Nathan, Matti Suominen & Joni Tasa (2026). The Intramonth Momentum Cycle.

18. Nikhil Devanathan, Dylan Rueter, Stephen Boyd, Emmanuel Candès, Trevor Hastie, Mykel J. Kochenderfer, Arpit Apoorv, David Soronow & Igor Zamkovsky (2026). Single-Asset Adaptive Leveraged Volatility Control. https://arxiv.org/abs/2603.01298

19. Tony Cooper (2010). Alpha Generation and Risk Smoothing using Managed Volatility. https://ssrn.com/abstract=1664823

20. Campbell R. Harvey, Yan Liu & Heqing Zhu (2015). … and the Cross-Section of Expected Returns. https://ssrn.com/abstract=2249314

21. David H. Bailey, Jonathan M. Borwein, Marcos López de Prado & Qiji Jim Zhu (2014). Pseudo-Mathematics and Financial Charlatanism: The Effects of Backtest Overfitting on Out-of-Sample Performance. https://ssrn.com/abstract=2308659

22. Paul Beekhuizen & Winfried G. Hallerbach (2015). Uncovering Trend Rules. https://ssrn.com/abstract=2604942

23. Alban Etienne, Jean-Jacques Ohana, Eric Benhamou, Béatrice Guez, Ethan Setrouk & Thomas Jacquot (2025). Revisiting the Structure of Trend Premia: When Diversification Hides Redundancy. https://arxiv.org/abs/2510.23150

24. Alberto G. Rossi, Yingguang Zhang & Yandi Zhu (2025). Short-Term Basis Reversal.

25. Daniel Gatto (2026). Adverse Selection Consumes the Touch: A True-Aggressor-Signed Maker-P&L Decomposition Across 13 CME Futures. https://github.com/DaruFinance/quant-mm-simulator

26. Tom Burdorf (2025). Learning to Rank: Enhancing Momentum Strategies Across Asset Classes.

27. Y.-F Li & G. Papanicolaou (2024). Statistical Arbitrage in Rank Space. https://arxiv.org/abs/2410.06568

28. Travis R. A. Sapp (2011). The 52-Week High, Momentum, and Predicting Mutual Fund Returns.

29. Ahmet Duran & Michael J. Bommarito (2009). A profitable trading and risk management strategy despite transaction cost. https://ssrn.com/abstract=1509811

30. Michael C. Münnix, Takashi Shimada, Rudi Schäfer, Francois Leyvraz, Thomas H. Seligman, Thomas Guhr & H. Eugene Stanley (2012). Identifying States of a Financial Market. https://doi.org/10.1038/srep00644

31. Martyn Tinsley (2026). Walk Forward Correlation: A Diagnostic for Over-Fitting and Structural Edge in Trading Strategy Optimisation.

32. Michael Cook, Edward Hoyle, Matthew Sargaison, Dan Taylor & Otto Van Hemert (2017). The Best Strategies for the Worst Crises. https://ssrn.com/abstract=2986753

33. Ákos Maróy (2025). Improvements to Intraday Momentum Strategies Using Parameter Optimization and Different Exit Strategies.

34. Trinh Le & Ummul Ruthbah (2023). Trend-following Strategies for Crypto Investors. https://ssrn.com/abstract=4551518

35. Roberto Croce, Avishek Hazrachoudhury, Sladja Carton & Suren Karapetyan (2026). Trend-following crisis alpha: Does it come from beta timing or market selection?.