WorksEconomics, Econometrics and Finance
Hard to Fool: A Bounded LLM Research Loop for Quantitative Finance, and What Two Dozen Campaigns Found
Abstract
Agentic research loops that let a language model propose a change, run a fixed experiment and keep whatever improves the metric work well in machine learning. In trading research the same loop is a machine for producing false discoveries: a backtest that "improves the metric" is, by default, overfit, look-ahead, a multiple-comparisons artifact, cost-free, or market drift wearing a strategy. We describe Autonomous Quant Researcher, an open-source loop built on the opposite premise: the model proposes, and only trusted code decides. Proposals are strict declarative JSON over a fixed vocabulary, so there is no model-written code to sandbox. Evaluation runs on immutable chronological stores split into discovery, validation and a final holdout that can be opened exactly once, after a durable receipt is written. Acceptance is a conjunction of absolute gates (positive edge and net P&L after costs, a positive weekly-cluster bootstrap lower bound, survival at elevated costs, per-year stability and sample floors), never a single score. Duplicates are refused before they spend a trial, and every formulation, refusal, crash and result is kept in a hash-chained, append-only ledger. We report every campaign run to date, across 0DTE options priced from traded option bars, overnight SPX verticals, daily ETF rules, technical-analysis premises, Kalshi crypto binaries, earnings options, off-exchange volume and a pre-registered alternative-data campaign. Most hypotheses fail, and we report them. Two results survive their pre-registered bars. The clearest example of the loop working is a failure: a language-model survivor that passed discovery and validation with Sharpe ratios near 3 and fell to 0.26 on its single holdout use. The value of such a loop lies less in what it finds than in how hard it is to make it report something that is not there.
Introduction
Two open projects released in 2026 made a simple observation about agents: the agent is the cheap part, and the loop around it is the product. Karpathy’s autoresearch [9] lets an agent edit a single training script, run a fixed five-minute job, keep the change if the validation metric improved and discard it otherwise, all night. Greyling’s Loop Engineering [5] turns the loop itself into a first-class artifact: a persistent contract, explicit budgets, deterministic gates, an audit trail and maker/checker separation.
Finance is a domain where that contract breaks in a specific way. The problem of data snooping in asset pricing is old [12, 16], with formal corrections for searching many rules at once [6, 24], and the evidence that most published return predictors do not survive correction for the number of tests tried is now extensive [7, 8]. Backtest overfitting is easy to produce and hard to detect after the fact [4, 2, 3], and published predictability decays once it is known [13]. A language model that proposes thousands of hypotheses per night does not repair any of this. It industrializes it.
Recent LLM-agent systems for trading and alpha mining, such as TradingAgents [17], FinMem [29], AlphaAgent [15], R&D-Agent-Quant [11], tree-search factor mining [25] and code-based alpha evolution [23], and role-based agents for equity portfolio construction [27], focus on making the proposer better: richer roles, memory, reflection, factor generation and reinforcement. Others ask what a language model can extract in the first place, for example whether it tracks evolving signals in corporate disclosures [28]. Work on the evaluation side is newer. Deep et al. [21] validate interpretable hypotheses walk-forward across many out-of-sample periods with costs; Li et al. [22] argue that historical backtests let a language model “time travel” through outcomes it has already seen, and benchmark live instead; and a scoping review by Yao, Zheng and Li [26] finds execution realism weakly standardized across LLM trading studies, which limits comparison and reproduction. Closest to our aim, Gençay [20] builds leakage safety and search-aware deflation into a strategy-discovery system and finds that no LLM-discovered strategy survives certification at the agent’s own trial count; a planted look-ahead oracle with a Sharpe ratio of 35 passes both the deflated Sharpe ratio and the probability of backtest overfitting, so statistical correction cannot stand in for structural leakage control. Chen and Pu [19] remove look-ahead at inference by logging an agentic model’s daily stock scores in real time. Evidence from production points the same way: across two live fleets of LLM trading agents, Barton et al. [18] find that the operating layer (sizing constraints, rendered choices, order mechanics) shapes behaviour more than the choice of model, and that neither fleet shows a directional edge. We share Gençay’s verdict and differ in the instrument: hypotheses are declared as data and pre-registered, the holdout can be spent once, and every campaign, across options, microstructure, prediction markets and equities, is reported with its rejections. We focus on that other half of the loop. Our contribution is a reference design, and an open implementation, for the part that decides whether a proposal is true, together with a complete report of what it decided.
Concretely, this paper contributes:
A declarative hypothesis boundary: the model emits data, not code, so the attack surface and the space of searchable hypotheses are both fixed in advance (§ 2.1).
A staged evidence lifecycle with a one-use holdout enforced by a durable receipt, admission control that refuses duplicates before they spend a trial, and acceptance by a conjunction of absolute gates (§ 2.2).
A tamper-evident ledger and portable findings that carry their own provenance and authority (§ 2.4).
A complete campaign report, including every rejection (§ 3).
A list of failure modes that the gates did not catch the first time, and the change each one forced (§ 4).
The engine is open source [1], with a reference mission and a public pre-registration whose ledger can be verified independently. Market-data files are not redistributed.
Design
The declarative hypothesis boundary
In the current schema (version 3) a proposal has exactly two fields, hypothesis and rationale. The hypothesis contains a schema version, an entry minute on the five-minute regular-session grid, a bounded Boolean signal expression over approved causal observations (return from the open, return between two earlier times, first touch before entry) and exactly one approved defined-risk options structure (a call debit spread or a put credit spread) with bounded dimensions. Trusted functions interpret the declaration. A response with extra fields, invalid JSON, unsupported vocabulary, malformed timing, excessive nesting or out-of-range parameters is rejected before any scientific evaluation.
This boundary replaces an earlier design, preserved in the repository, in which the model returned an executable Python candidate. That design needed a sandbox (credential stripping, import and capability denial, a content-hashed evaluator) and it still left the candidate free to express things the evaluator did not anticipate. Removing the executable field removes the problem rather than containing it: there is nothing to sandbox, and every admissible hypothesis is, by construction, one the evaluator knows how to price.
Staged evidence and the one-use holdout
Stage stores are materialized together from one historical source, but each access capability resolves only its own stage, and every store carries a content hash and a manifest hash. Each campaign freezes its own windows. In the reference SPY mission on daily and modeled data, discovery is 2016–2021, validation 2022–2024 and the final holdout 2025 onward. The 0DTE campaigns priced from traded option bars start when those bars do: discovery is 2024-01-19 to 2025-08-05, validation 2025-08-06 to 2026-02-05 and the holdout 2026-02-06 to 2026-08-10.
Freeze before the first result. The mission, policy, vocabulary, interpreter and evaluator versions, stage manifests, base and elevated costs, trial budget, diversity rules and gates are hashed into a campaign contract before the first proposal is evaluated. Changing any of them creates a new campaign; a non-terminal campaign cannot continue under changed semantics.
Admission. Formulations and scientific trials are counted separately. Invalid documents are recorded but do not spend a trial. Semantic duplicates and near-duplicates (one numeric knob turned on the same idea) are refused, and a family quota admits at most two candidates per descriptor family. A 50-trial budget therefore buys 50 distinct ideas, and the trial count reported with any result is the number that was actually tried.
Shortlist. After the discovery budget closes, only candidates that passed every discovery gate are eligible. The coordinator ranks them by their weakest gate component, prefers distinct clusters, freezes at most three and hashes the shortlist. Validation evaluates only that set and divides the 5% lower-tail probability by the shortlist size.
The holdout. The ordinary worker stops at holdout_ready and cannot open the final stage. Holdout access is resolved only after an irreversible receipt has been written durably; a crash after the receipt still consumes the holdout and cannot be retried. Budget exhaustion is a pause, never a success.
Gates, not scores
Every stage must pass every applicable gate:
absolute edge after base costs ;
base-cost mean net P&L ;
weekly-cluster bootstrap lower bound on net P&L (2,000 deterministic resamples, clustered by ISO week; cf. block bootstrap [10]);
mean net P&L at elevated costs ;
finite metrics and priced rows;
sample floors. In the reference SPY mission these are 120 trades and 40 effective weeks in discovery; 60 and 24 in validation, with evidence in each of 2022, 2023 and 2024, at least 15 trades and 8 weeks per year, and no negative year; and 30 and 12 in the holdout.
Floors scale with a campaign’s windows and are frozen in its contract. The traded-bar 0DTE campaigns, with shorter windows, require 40, 15 and 15 trades. A ranking score, the minimum of absolute edge, the bootstrap bound, elevated-cost P&L and the worst year, chooses among candidates that passed; it cannot admit one. A control-relative metric is computed but is diagnostic only, so a weak control cannot rescue a candidate.
Modeled option prices are not market prices. Actual OPRA evidence is a separate typed gate with its own coverage floors and states (unavailable, insufficient_coverage, failed_evidence, passed); replay consumes only a hash-verified quote artifact with no fallback price source.
Ledger, provenance and authority
Every event (formulation, admission decision, trial, crash, stage transition, holdout receipt) is appended to a ledger in which each record carries the hash of its predecessor. Editing, removing or reordering an event makes the campaign fail closed. A surviving finding is exported as portable JSON built only from frozen aggregate evidence: contract and manifest hashes, per-stage evidence, limitations and invalidation conditions, and an immutable authority stamp (research_only, execution_authorized: false). Whatever consumes a finding downstream knows exactly what it is holding.
The same machinery supports pre-registration for campaigns whose data does not yet exist: a specification and the code that will compute its verdict are bound by SHA-256 into the ledger, typed data requirements refuse to run until met, and each campaign is judged once, with the receipt written before any data is read [14].
Model independence
The evaluator is deterministic CPU code; only the proposer needs a model. Any OpenAI-compatible endpoint can propose. The reference configuration is a locally served 7B-parameter coding model. Because the evaluator does not depend on who proposes, changing the model cannot move a metric, only the candidates.
Results
Table 1 lists every campaign recorded in the repository as of this writing. Each campaign has a record holding its pre-registration (where one exists), result tables and limitations. Some campaigns predate schema 3 and ran through the executable-candidate loop or a fixed sweep harness under the same staged lifecycle, with the splits, costs and floors frozen in each campaign’s own files; one (Kalshi maker side) is explicitly labelled hypothesis generation, because its data had already been read. We report all of them.
| Campaign | Question and scale | Verdict |
| 0DTE and overnight index options (actual traded option bars unless noted) | ||
| spy-odte-credit | SPY 0DTE put credit, w5/o5, 629 sessions | passes both stages at 0.5% cost; fails drift-neutral bound |
| geometry cross | 5 pre-registered geometries | 0 of 5; best (w5/o2) 7 of 8 conditions |
| QQQ replication | one named cell, 635 sessions | 7 of 8; same failing condition |
| calm conditioning | prior-day stress filter | refuted; mechanism inverted |
| odte-intervention | 120 management rules vs. hold | 0 supported |
| qqq-0dte-holdout | LLM survivor, holdout spent once | rejected: Sharpe 2.94 0.26 |
| side-asymmetry | post-hoc call/put asymmetry, 8 new markets | replication refuted |
| overnight-spx | 40 LLM trials, modeled SPX verticals | 0 passed discovery |
| entry sweeps | pre-registered grids; w20/o15 promoted | holdout spent once, rejected |
| 1DTE horizon | 228 cells, one full validation pass | rejected as drift (56%) |
| odte-call-side, friday-odte-names | single-name 0DTE | not promoted |
| Daily bars and technical analysis | ||
| bars-rules | 232 single-rule cells, 8 ETFs | 4 discovery passers, 0 promotions |
| ta-premises | Fibonacci, channels, MA ribbon, support | all fail validation |
| ta-indicators | 8 indicator claims | 0 cleared discovery |
| smartfit-channels | 5 pivot-channel claims | 0 cleared discovery |
| hmm-regime | HMM regime system, SPY daily, BTC hourly | rejected on both |
| idio-gap | idiosyncratic gap continuation, 1,945 events | refused; crash-day reversion found |
| earnings-vol, pead | earnings straddles, wings, drift | no edge |
| Crypto and prediction markets | ||
| kalshi-btc-hourly | 138 conditional hypotheses | 0 cleared; one look-ahead artifact |
| kalshi-btc-15m | 15-minute binaries | not supported; fails replication |
| kalshi-maker-side | maker fills vs. adverse selection | no (hypothesis generation) |
| btc-sessions, btc-kalshi-windows | weekend and overnight BTC | 0 promotions |
| flow-crypto-etf-weekend | Friday ETF options flow, 84 cells | 0 cleared Bonferroni |
| Cross-section and alternative data | ||
| dpi-offexchange | off-exchange share vs. forward returns | clears pre-registered bar |
| gpu-leads-revisions | GPU rental prices revisions | primary waiting; proxy v3 fails |
Table 1. Every campaign run to date. “w5/o2” denotes a spread 5 points wide with the short strike 2 points out of the money. Costs and gates are those frozen in each campaign’s own files.
A language-model survivor meets its holdout
The most instructive single result is a QQQ 0DTE campaign in which the language-model proposer, in its thirteenth run, produced a rule that sold the near-money put credit spread on Mondays and Tuesdays when the credit exceeded 10% of width. On the traded-bar windows above, and under that campaign’s own floors (40, 15 and 15 trades), it passed every discovery and validation gate (Table 2). On explicit instruction the holdout was spent once. Eight of nine holdout gates passed; the rule stayed net positive, survived doubled costs and the removal of its best day. But expectancy fell by about 90%, the Sharpe ratio by an order of magnitude, and one half of the holdout window was net negative. The rule was rejected, and the holdout window is now permanently marked as spent for this candidate.
| Discovery | Validation | Holdout | |
| Net per trade (% of width) | +5.69 | +4.82 | +0.52 |
| Net Sharpe | 2.94 | 3.34 | 0.26 |
| Trades | 221 | 43 | 72 |
| Status | passed | passed | rejected |
Table 2. The QQQ language-model survivor across its three stages.
Two points matter for the design. First, the in-sample numbers were not marginal; a loop that accepts on a validation Sharpe above 3 would have promoted this rule. Second, the frozen specification for that campaign had been written to a temporary path and was not on disk when the holdout was requested. Rather than evaluate under a guessed specification, it was reconstructed and checked against the two already-spent stages (discovery Sharpe matched the recorded value to six decimals, and the dataset hash matched the staging manifest) before the holdout was read. The campaign contract is now persisted with the ledger, so this cannot recur.
The 0DTE put-credit premium: real, and not enough
The unconditional SPY 0DTE put credit spread (width 5, short strike 5 points out of the money, 10:00 ET entry, held to settlement) passed every discovery and validation gate at a round-trip cost of 0.5% of width, with +55.7 and +71.6 basis points of width per session respectively, and a gross win rate of 95.2%. Three follow-ups, each pre-registered as a new contract, then tested whether this was a strategy or a byproduct.
Drift. Regressing each session’s P&L on SPY’s own return over the same window attributes 13–30% of the payoff to market direction. The drift-neutral residual is positive in both splits (+74 and +100 bps of width gross) but its weekly bootstrap lower bound is marginally negative in both. By the rule written before the attribution ran, the structure is not promoted.
Geometry. Across five pre-registered geometries the premium is monotone in moneyness: offset 2 beats 5 beats 10, and offset 10 is negative. The best cell, 5/ω2, passes every condition except one, the discovery raw bootstrap bound, which misses by 1.35 bps of width. A pre-registered QQQ replication of the same cell in percentage terms fails exactly the same condition at a similar magnitude. The worst discovery weeks on both tapes are the same calendar shocks (April 2024, August–September 2024, March–April 2025).
Conditioning. A filter that skips entry after a prior-day move of at least 1% (chosen by convention, not fit) makes every bootstrap worse. The losses come from the first day of a shock, which follows a calm day and is not forecastable from prior-day information; by the next day volatility has repriced and selling pays best.
The honest summary is that the near-money 0DTE put premium is real, has the same shape on two underlyings, is mostly not drift, and earns its return by absorbing unforecastable first-day shocks. The pre-registered bar, which asks for a positive fifth percentile of weekly means across a window containing three such shocks, is not met, and the bar was not moved. The breakeven entry cost for full gate compliance lies between 0.5% and 1.0% of width, which turns an open question about the strategy into a measurable question about execution.
What the TA and daily-rule campaigns show
In the 232-cell daily-rule grid, all four discovery passers used the same feature in both directions: long after a 20-day decline and long after a 20-day advance. A condition whose opposite passes equally carries no information; what passed was the long side in a window where holding these ETFs paid. The gates caught sample and stability problems but could not by themselves detect that the whole discovery window pointed one way. All three shortlisted cells failed every performance gate in validation. In the technical-analysis campaigns, one premise (a regression channel) passed all nine discovery gates and then lost 22 bps per session in validation; every premise also underperformed simply being long.
What survived
Two results clear their own pre-registered bars and are the subject of separate papers. The share of a stock’s volume executed off-exchange, built from public FINRA data, has a negative information coefficient at 1, 5 and 20 days ( between −3.1 and −5.8), negative in every calendar year 2021–2026, concentrated in the top decile, and robust to size and short-term reversal; the vendor-published “dark pool index” predicts nothing. In the idiosyncratic-gap campaign the pre-registered continuation hypothesis was refused, while reversion of the day’s losers on market-wide crash days survived every gate. We report both as research findings with their limitations, not as strategies.
Failure modes the gates missed, and the change each forced
Drift as alpha. An unconditional long-delta credit spread collects the market’s upward drift whether or not anything is mispriced. An overnight SPX 1DTE at-the-money call debit spread passed every validation gate with a weekly Sharpe ratio of 2.27; 56% of its validation payoff was the underlying’s own move, and the bootstrap bound on the residual was negative. A pre-registered drift control (regress each trade on the underlying’s move over the same hold, gate the residual) is now part of the bar for this family, the drift decomposition is carried in the snapshot itself, and benchmark-relative payoffs are available to every campaign.
A one-direction discovery window. Gates evaluated on one window cannot see that the window had one direction. The rule recorded with the grid is that a condition is read together with its opposite, and a passing pair is a long-side artifact, not a signal.
Costs assumed, not measured. Two candidates were killed by a doubled-cost gate against a cost nobody had measured, and one holdout was spent before its cost premise was measured. The overnight SPX promotion was spent under time pressure on a single night’s cost capture and came back rejected. Campaigns now report the breakeven cost alongside the result, and the written rule is that measured entry costs precede any holdout spend.
Look-ahead through a feature. The single formal pass among 138 Kalshi hypotheses conditioned on volume in the same window it predicted. Cells that condition on in-window information are now marked descriptive-only and cannot count as evidence.
A lost specification. See the holdout case above; the contract is now persisted with the ledger.
Spending evidence for nothing. When every premise was already negative in validation, the TA holdout was left unread: no outcome could change, so reading it would only destroy a clean window.
Limitations
The loop makes it hard to report something that is not there; it does not make true things easy to find. Its vocabulary is deliberately small, and a hypothesis outside it cannot be tested until trusted code is extended. Several campaigns use short samples (for example 46 days of Kalshi data, 60 weekends of ETF flow), and their nulls are statements about power as much as about markets. Option results from 2024 onward cover no bear market of 2022’s kind. Some modeled prices remain where actual quotes were not available; those campaigns say so. Costs in several campaigns are assumptions with stated breakevens rather than measurements. Finally, this paper is written by the system’s author; the full campaign records behind each number are available from the author on request.
Conclusion
A research loop in quantitative finance should be judged first by how hard it is to fool and only second by what it finds. Making the model propose data instead of code, separating stages physically, letting the holdout be spent once, accepting only on a conjunction of absolute gates, refusing duplicates before they cost a trial and keeping every failure in a chained ledger together produce a system whose most common output is a well-documented “no”. We think that is the correct base rate, and that publishing it is the point.
Reproducibility. Engine code, a reference mission and the public pre-registration with its hash-chained ledger are at https://github.com/henryzhangpku/autonomous-quant-researcher. The test suite runs offline without keys. Campaign records are available from the author on request. A browser demonstration runs the engine under Pyodide on a synthetic tape with a planted regime that discovery finds and validation refutes.
About the author. Henry Zhang is a quantitative engineer with fifteen years building research platforms and production trading systems, including eight years at BlackRock and two at PIMCO. The research engine used here is open source at https://github.com/henryzhangpku/autonomous-quant-researcher; more research and notes at https://henryzhang.substack.com and https://www.officialhenryzhang.com.
Disclaimer. All backtests are hypothetical research evidence and not investment advice.
References
- [1]H. Zhang. Autonomous Quant Researcher. https://github.com/henryzhangpku/autonomous-quant-researcher, 2026.
- [2]D. H. Bailey and M. López de Prado. The deflated Sharpe ratio: correcting for selection bias, backtest overfitting, and non-normality. Journal of Portfolio Management, 40(5):94–107, 2014.
- [3]D. H. Bailey, J. M. Borwein, M. López de Prado, and Q. J. Zhu. The probability of backtest overfitting. Journal of Computational Finance, 20(4):39–69, 2017.DOI
- [4]D. H. Bailey, J. M. Borwein, M. López de Prado, and Q. J. Zhu. Pseudo-mathematics and financial charlatanism: the effects of backtest overfitting on out-of-sample performance. Notices of the American Mathematical Society, 61(5):458–471, 2014.DOI
- [5]C. Greyling. Loop Engineering. https://github.com/cobusgreyling/loop-engineering, 2026.
- [6]P. R. Hansen. A test for superior predictive ability. Journal of Business & Economic Statistics, 23(4):365–380, 2005.
- [7]C. R. Harvey, Y. Liu, and H. Zhu. . . . and the cross-section of expected returns. Review of Financial Studies, 29(1):5–68, 2016.
- [8]K. Hou, C. Xue, and L. Zhang. Replicating anomalies. Review of Financial Studies, 33(5):2019–2133, 2020.
- [9]A. Karpathy. autoresearch. https://github.com/karpathy/autoresearch, 2026.
- [10]H. R. Künsch. The jackknife and the bootstrap for general stationary observations. Annals of Statistics, 17(3):1217–1241, 1989.DOI
- [11]Y. Li et al. R&D-Agent-Quant: a multi-agent framework for data-centric factors and model joint optimization. arXiv:2505.15155, 2025.arxiv.org/abs/2505.15155
- [12]A. W. Lo and A. C. MacKinlay. Data-snooping biases in tests of financial asset pricing models. Review of Financial Studies, 3(3):431–467, 1990.
- [13]R. D. McLean and J. Pontiff. Does academic research destroy stock return predictability? Journal of Finance, 71(1):5–32, 2016.
- [14]B. A. Nosek, C. R. Ebersole, A. C. DeHaven, and D. T. Mellor. The preregistration revolution. Proceedings of the National Academy of Sciences, 115(11):2600–2606, 2018.
- [15]Z. Tang et al. AlphaAgent: LLM-driven alpha mining with regularized exploration to counteract alpha decay. arXiv:2502.16789, 2025.arxiv.org/abs/2502.16789
- [16]H. White. A reality check for data snooping. Econometrica, 68(5):1097–1126, 2000.DOI
- [17]Y. Xiao, E. Sun, D. Luo, and W. Wang. TradingAgents: multi-agents LLM financial trading framework. arXiv:2412.20138, 2024.arxiv.org/abs/2412.20138
- [18]T. J. Barton, C. Constantakis, P. Hauseman, A. Mous, A. Hoffman, B. Bergeron, and H. Goodreau. What LLM trading agents actually do in production: a six-month, population-scale record from two fleets. arXiv:2609.05663, 2026.arxiv.org/abs/2609.05663
- [19]Z. Chen and D. Pu. Autonomous market intelligence: agentic AI nowcasting predicts stock returns. arXiv:2601.11958, 2026.arxiv.org/abs/2601.11958
- [20]E. Gençay. What survives honest evaluation? Leakage-safe, search-aware assessment of LLM-driven trading strategy discovery. arXiv:2608.27734, 2026.arxiv.org/abs/2608.27734
- [21]G. Deep, A. Deep, and W. Lamptey. Interpretable hypothesis-driven trading: a rigorous walk-forward validation framework for market microstructure signals. arXiv:2512.12924, 2025.arxiv.org/abs/2512.12924
- [22]C. Li, Y. Shi, et al. Time travel is cheating: going live with DeepFund for real-time fund investment benchmarking. arXiv:2505.11065, 2025.arxiv.org/abs/2505.11065
- [23]F. Liu, Y. Huang, et al. Cognitive alpha mining via LLM-driven code-based evolution. arXiv:2511.18850, 2025.arxiv.org/abs/2511.18850
- [24]J. P. Romano and M. Wolf. Stepwise multiple testing as formalized data snooping. Econometrica, 73(4):1237–1282, 2005.
- [25]Y. Shi, Y. Duan, and J. Li. Navigating the alpha jungle: an LLM-powered MCTS framework for formulaic factor mining. arXiv:2505.11122, 2025.arxiv.org/abs/2505.11122
- [26]J. Yao, Z. Zheng, and B. Li. Execution realism and reproducibility in LLM-based trading systems: a systematic scoping review and evidence audit. arXiv:2606.08285, 2026.arxiv.org/abs/2606.08285
- [27]T. Zhao, J. Lyu, S. Jones, H. Garber, S. Pasquali, and D. Mehta. AlphaAgents: large language model based multi-agents for equity portfolio constructions. arXiv:2508.11152, 2025.arxiv.org/abs/2508.11152
- [28]C. Choi, Y. Kim, Y. Yu, Y. Cha, V. Z. Golkhous, I. Halperin, et al. From text to alpha: can LLMs track evolving signals in corporate disclosures? arXiv:2510.03195, 2025.arxiv.org/abs/2510.03195
- [29]Y. Yu et al. FinMem: a performance-enhanced LLM trading agent with layered memory and character design. arXiv:2311.13743, 2023.arxiv.org/abs/2311.13743