PaperlayerView exact revision on arXiv

Source-linked AI summary

Agentic Empirical Asset Pricing: Methodological Foundations

Yingjian Pan, Xiaowei Ding, Kay Giesecke

arXiv:2609.00731v1cs.AIcs.LGq-fin.ST

TL;DR

AEAP lacks a shared definition and evaluation discipline for autonomous discovery systems, limiting comparison and replication. The paper defines AEAP, proposes multi-axis and rolling evaluation, and finds that no system excels on every standard while residual backbone leakage remains.

  • Problem

    AEAP lacks a shared definition and standard for evaluating autonomous discovery processes rather than only their outputs, limiting comparison, replication, and cumulative research.

  • Method

    The paper defines AEAP, establishes point-in-time multi-axis and rolling out-of-sample evaluation, and applies the framework to SEADS and five baselines.

  • Results

    No system excels on every standard across both panels, and even multi-axis scoring cannot settle the best system from one decision date alone.

  • Takeaways & Limitations

    AEAP systems should be assessed jointly across multiple output dimensions and through historical re-execution of the discovery process.

  • Takeaways & Limitations

    The backbone has residual soft information leakage because one current LLM is used across historical decision dates rather than release-dated models.

Abstract

from arXiv · show

Recent advances in LLM agents enable a new paradigm for asset pricing, which we call Agentic Empirical Asset Pricing (AEAP): systems that autonomously conduct the scientific discovery process itself. We define AEAP and identify its core building blocks. Existing evaluation practices backtest only the outputs (factors or trades), not the autonomous discovery system that produced them. We focus on factor discovery, contributing a reference architecture, a rigorous evaluation standard for discovered factors, and a method for out-of-sample backtesting the discovery system. As a concrete instance of that architecture, we evaluate SEADS against five re-implemented baselines on two US equity panels using this standard: no single metric ranks the systems consistently, motivating evaluation on multiple axes at once. A separate rolling re-execution then asks the complementary question of whether the discovery process itself, not one static output, is reliable. We also report negative findings and limitations that surface further evaluation pitfalls for future AEAP systems.

1 Introduction

Agentic Empirical Asset Pricing (AEAP) describes LLM agents that autonomously generate hypotheses, implement code, and statistically validate discoveries. The paper establishes shared foundations and proposes an evaluation methodology for factor-discovery systems that assesses both reported outputs and the discovery process.

  • Motivation: AEAP shifts empirical asset-pricing research from human-led discovery toward autonomous hypothesis generation, coding, and statistical validation.
  • Motivation: Existing AEAP evaluations backtest factors or trades but overlook how autonomous systems can obtain lucky discoveries through their research processes.
  • Contributions: The paper defines AEAP, identifies five common building blocks, and distinguishes the paradigm from classical machine learning and human-in-the-loop LLM copilots.
  • Contributions: For factor discovery, it proposes a reference architecture and a joint evaluation standard based on productivity, performance, and novelty, plus repeated historical re-execution.
  • Contributions: The paper instantiates this framework in SEADS and evaluates it alongside five re-implemented baselines on two comprehensive US equity panels.

2 Agentic Empirical Asset Pricing

Agentic Empirical Asset Pricing (AEAP) is a proposed third wave of empirical asset pricing in which autonomous LLM-based agents pursue scientific-discovery objectives with minimal human intervention, while invoking econometrics and machine learning within their loops [Gu et al., 2020, Kozak et al., 2020]. An AEAP system autonomously completes a full hypothesis–formalization–evaluation cycle over asset-pricing data, motivating evaluation of both its inner discovery cycle and repeated outer-loop execution.

  • Motivation: AEAP extends traditional econometrics and machine learning into a third wave where autonomous LLM-based agents pursue asset-pricing objectives with minimal human intervention [Gu et al., 2020, Kozak et al., 2020].The objectives include testing theories, discovering factors, explaining expected returns, forecasting, studying anomalies, evaluating portfolios, and estimating risk premia.
  • Definition: AEAP requires LLM-based agents to autonomously execute at least one complete hypothesis, formalization, and evaluation cycle over asset-pricing data.Human researchers cannot perform any of those three steps for the system to qualify.
  • Architecture: Across AEAP systems, five recurring building blocks are hypothesis generation, formalization into a checkable artifact, execution or estimation, evaluation, and iterative scientific discovery.The supplied passage explicitly introduces these recurring components, though its excerpt ends during the execution-or-estimation description.
  • Scope: AEAP spans factor discovery, trading-policy evaluation, and structural-model testing, but systems pursuing different objectives are not directly comparable to factor-discovery systems.The paper focuses on factor discovery because it has the most existing systems and the best-defined statistical target for testing the discovery process.
  • Evaluation framework: Sound AEAP evaluation must enforce information available at each decision date across two nested loops: the inner discovery cycle and an outer loop that re-runs discovery across historical dates.This differs from scoring one static output on a single split and motivates the paper’s focus on evaluating the discovery system itself.

3 Factor Discovery

This section formalizes factor discovery as finding an out-of-sample-valid, non-redundant factor set and instantiates that objective in SEADS. It defines multi-axis evaluation of discovered outputs and rolling re-execution to assess the discovery process’s adaptivity.

  • Factor-discovery objective: The discovery target is a factor set that is simultaneously out-of-sample valid and non-redundant relative to an explicit, panel-specific reference set.No single metric captures both properties; admissibility requires correlations with every reference factor to remain below a threshold.
  • Reference architecture: Five reimplemented baselines instantiate combinations of generators, sandboxed point-in-time executors, validation gates, and persistence mechanisms, with no common vocabulary.The executor and gate are intended to be schema-agnostic, while each baseline implements only a strict subset of the architecture’s building blocks.
  • SEADS implementation: SEADS separates planning from proposal, expands structured explore-or-exploit seeds, and executes candidate formulas through deterministic structural and point-in-time safeguards.Its sandbox restricts code to statically read columns, applies a look-ahead firewall, and supports up to two LLM-assisted repairs; the pipeline runs across both panels without hardcoded column names.
  • Validation and persistence: Candidates pass an ordered statistical gate requiring duplicate and self-collapse checks, coverage, Rank-IC t-stat, out-of-decade regime consistency, and robustness checks.Near-passes are self-evolved through the same gate using stage-specific feedback, while memory tiers preserve statistics, candidate outcomes, lessons, and admitted factors across rounds.
  • Evaluation standard: Systems are scored jointly on productivity, mean per-factor OOS Sharpe, and novelty, because performance, productivity, and novelty are non-substitutable properties of the reported factor pool.Productivity counts total admissions from a shared 300-candidate budget, while novelty uses maximum correlation with the system’s other admissions and the panel reference set.
  • Adaptive evaluation: Adaptivity is evaluated by rolling re-execution of the entire discovery loop at historical decision dates using only preceding sliding windows and persistent state.The rolling protocol carries forward the growing library, updates the novelty reference set with new admissions, and preserves lessons, distinguishing adaptive rediscovery from repeated cold static runs.

4 Experimental Setup

The experiments compare factor-discovery systems on two complementary US equity panels under matched, out-of-sample-only evaluation. They use re-implemented baselines and a locked validation gate to support fair method comparisons and system-level backtesting.

  • Data panels: Two US equity panels test novelty against either a heavily mined characteristics database or primitive, uncurated accounting and market variables.Panel A uses JKP’s ex-microcap US database, while Panel B uses CRSP/Compustat features and an external reference.
  • Evaluation design: Matched-budget, out-of-sample-only comparisons score productivity, performance, novelty, and spanning tests identically across panels, with lower values preferred only for novelty.Table 1 reports means across five replicates, with per-entry cross-replicate standard errors provided separately.
  • Data panels: Both panels use 2010–2019 in-sample data, with out-of-sample windows through December 2025 for Panel A and November 2024 for Panel B.The windows contain approximately 71 and 59 months, respectively, reflecting source-data availability.
  • Baselines: Five baselines—AlphaAgent, AlphaForge, RD-Agent(Q), QuantaAlpha, and Beyond Prompting—are re-implemented to isolate methods from differences in native-market data.Reported results are the authors’ reimplementations rather than the named systems’ published numbers; porting fidelity is not uniform [Tang et al., 2025, Shi et al., 2025, Li et al., 2025c, Han et al., 2026, Huang and Fan, 2026].
  • Validation and protocol: SEADS applies one locked internal validation gate across both panels before the backtesting-the-system protocol.The gate combines coverage, Rank-IC, out-of-decade regime, novelty, partial-IC, and Fama–MacBeth thresholds, including a fixed novelty cutoff and an ex-ante Bonferroni-dynamic Rank-IC rule.

5 Results

Results show that discovery systems materially outperform a raw-feature-only baseline, while rankings vary across panels and metrics. SEADS’s self-evolution mechanism is especially important, although removing the Planner creates a distinct over-admission failure mode.

  • Results: Mean Sharpe reaches 0.82–0.96 for all five discovery systems versus 0.62 for Raw 400 only; SEADS nominally leads at 0.96±0.04, but is not reliably distinguishable from Beyond Prompting or QuantaAlpha.Figure 2 jointly regresses admitted factors with the full raw characteristic panel, so the comparison measures value beyond what the raw panel already supplies.
  • Mechanism ablations: Removing Self-evolution reduces admits from 14.0 to 7.8 and Sharpe from 0.25 to 0.08, a 68% decline and the largest ablation drop.Removing the Planner instead increases admissions to 37.2 while reducing Sharpe to 0.10, indicating overproduction of weak factors rather than underproduction.

6 Limitations

The study’s conclusions are limited by small samples, dataset-dependent Sharpe, and several design choices that affect comparability, leakage control, and portfolio validity. Reported Sharpe is also gross of trading costs, while the admission gate certifies only marginal in-sample signals rather than joint out-of-sample portfolio contributions.

  • Sample size and power: Small samples and short OOS windows make most cross-system differences statistically unreliable, while replicate counts reflect API budget rather than deliberate power limits.Admitted-factor counts are in the tens per replicate, OOS windows span 59–71 months, and there are 5 static versus 3 rolling replicates.
  • Portfolio construction: Independent admission certifies marginal in-sample signals but does not guarantee coherent joint, non-monotone out-of-sample contributions in the combined portfolio.Later-admitted factors can interact with earlier ones in ways the admission gate does not certify.
  • Dataset dependence: Absolute Sharpe is not informative across panels because achievable Sharpe depends on dataset-specific alpha headroom.The study therefore compares systems within each panel rather than across panels using absolute Sharpe.
  • Information leakage: Backbone information leakage remains soft because one current GPT-5.2 model serves every role, although proposing from feature descriptions rather than realized returns keeps leakage near the safe end [Li et al., 2025b, Glasserman and Lin, 2024, Sarkar and Vafa, 2024, Lopez-Lira et al., 2025].General knowledge of historically successful factor families could still bias hypotheses, especially across the rolling protocol.
  • Look-ahead control: The look-ahead firewall may be overly conservative because static pattern matching can reject safe formulas alongside unsafe ones.A less conservative alternative exists but postdates these results.
  • Compute matching: Removing self-evolution retries and repairs collapses Sharpe by 68%, showing that matched candidate counts do not imply matched realized compute per admission.The 300-candidate budget matches initial proposals, but mutation retries and repairs add LLM calls beyond one proposal per slot.
  • Trading costs: Reported Sharpe is gross of trading costs, so transaction costs and turnover could disproportionately erode performance for frequent reranking or complex formulas.No reported number reflects this implementation gap.

7 Conclusion

The conclusion defines AEAP and presents a multi-axis standard for scoring discovered factors alongside historical, point-in-time backtesting of the discovery process. Across two panels, no system excels on every score, while negative findings show the field remains unsettled.

  • Contributions: The paper defines AEAP and identifies its building blocks for systems that discover their own hypotheses rather than test a researcher’s.This addresses the absence of a shared definition for autonomous empirical discovery.
  • Evaluation framework: It proposes a multi-axis output-facing standard for scoring discovered sets jointly and a method for backtesting discovery systems across historical decision dates.The approach evaluates the process, not only the output of one run.
  • Contributions: SEADS instantiates the framework and is evaluated against five re-implemented baselines.The comparison motivates evaluating AEAP systems across multiple axes rather than relying on a single score.
  • Empirical findings: Across both panels, no system excels on every standard score, and one decision date cannot identify the best system.Historical re-execution complements static scoring by testing whether the discovery process holds up under point-in-time constraints; negative findings include non-monotone combination and a soft backbone-leakage residual.

Ethical Statement · Appendix

The ethical statement frames automated trading-signal discovery as dual-use, with overstated performance claims and deployment-related crowding as primary concerns. The appendix provides full methodology and results supporting the main paper.

  • Ethical Statement: Automated trading-signal discovery is characterized as a dual-use capability whose primary risk is overstated performance claims, rather than one dangerous capability.
  • Ethical Statement: The evaluation standard is deliberately deflationary, aiming to reduce rather than inflate the field’s headline claims.
  • Ethical Statement: Out-of-sample scoring of discovery systems is presented as a mechanism for reducing inflated performance claims.
  • Ethical Statement: Reporting negative findings is presented as another mechanism for restraining inflated claims in automated discovery research.
  • Ethical Statement: Automated discovery at scale may contribute to market crowding if many actors deploy identical systems; the evaluation panels include small-cap names.
  • Appendix: The appendix supplies full methodology and results supporting the main paper, with specific tables and figures cited where relevant.

A Data Construction Detail

The study uses two US equity data panels with explicit size, source, novelty-reference, and point-in-time conventions. All feature sources enforce a uniform one-month-ahead return definition to prevent silent downstream corruption.

  • Panel A: JKP characteristics: Panel A uses JKP global factor characteristics for US-listed stocks, excluding micro and nano stocks via monthly NYSE market-cap breakpoints without look-ahead.The selected variant retains small, large, and mega stocks; the pipeline contains four size-filtered variants.
  • Panel B: CRSP/Compustat base features: Panel B’s base-features panel contains 87 primitive CRSP/Compustat accounting and market variables across 13 categories and 1,213,574 stock-months.The category sizes range from 3 to 13 variables.
  • Panel B’s novelty reference: Panel B’s novelty reference is Bryzgalova et al. [2025]’s 45-characteristic “Missing Financial Data” panel, restricted to the prior-month market-equity top 80% without look-ahead.It covers 2,402,623 stock-months, 22,336 unique stocks, 1977–2024, and six categories.
  • Point-in-time convention: All feature sources store ret(t) as the lead-1-month return realized over (t, t+1], replacing track-specific lag parameters with a uniform contract.The paper treats this convention as an invariant because one incorrect return-column definition could corrupt every factor built on that panel.

B Reference Architecture Component Audit … G Per-Baseline Fidelity Notes

The paper specifies SEADS as a closed-loop factor-discovery system with deterministic validation, look-ahead protection, self-evolution, and reproducible combination. It also defines walk-forward evaluation of the discovery process and documents fidelity choices for each re-implemented baseline.

  • B Reference Architecture Component Audit: The reference-architecture audit compares five re-implemented systems across four building blocks plus system-level backtesting, omitting point-in-time execution because it is uniform.Table 4 treats AlphaForge as the standard non-agentic comparison baseline rather than evidence that AEAP building blocks are optional.
  • C The SEADS Discovery Loop: Pseudocode: SEADS’s loop generates, validates, mutates, and preserves candidate factors, while every surveyed system implements generation, evaluation, and cross-round persistence in some form.SEADS’s self-evolution is more elaborate, using four components for persistence versus one each in the other systems; near-misses are mutated only after clearing specified progress conditions.
  • D Validation Gate and Firewall Detail: Validation is deterministic: candidates must pass every configured check in order, with short-circuit rejection and a final sign backstop.The firewall’s rule classes and the gate’s exact execution order are specified separately; the gate includes a self-collapse check and sign agreement across realized metrics.
  • E Pipeline Internals: SEADS controls implementation risks through locked or sequential shared state, collision-safe naming, scoped prompts, structural planner backstops, and formula-trace grounding.Its Ridge combinator uses rank-normalized admitted factors with α = 1 to remain stable, comparable, and exactly reproducible across systems and panels.
  • F Backtesting-the-System: Full Walk-Forward Protocol: Walk-forward evaluation re-executes the discovery process at historical decision dates, supporting default revalidation without new LLM calls or full reproposal with fresh discovery.Reproposal and combinator-refit cadences are independent; the evaluation uses annual reproposal and four-month refit, holding the admitted library fixed between reproposals.
  • F Backtesting-the-System: Full Walk-Forward Protocol: The rolling protocol preserves SEADS’s library, novelty reference set, and lessons across reproposal events, whereas baselines re-invoke their own drivers at each new decision date.Both approaches share the same reproposal/refit tick structure, but only SEADS carries forward these cross-refresh mechanisms.
  • G Per-Baseline Fidelity Notes: Baseline ports disclose system-specific fidelity: some use released code, some require documented substitutes, and RD-Agent(Q) reports native, firewall-audited, and firewall-clean views separately.Beyond Prompting preserves published thresholds, including its Harvey–Liu–Zhu t>3.0 hurdle, while QuantaAlpha follows its paper’s stated admission rule when code differs.

H Formula-Complexity Case Study: Interactions vs. Additive Spreads · I Broader Related-Work Landscape · J Reproducibility Notes

The case study finds that SEADS discovers predominantly multiplicative, conditional interactions, whereas the baselines mainly produce additive spreads, which can be redundant in linear models. The paper situates AEAP within related factor-discovery and agentic-finance research and documents deterministic, leakage-resistant evaluation and reproducibility choices.

  • H Formula-Complexity Case Study: Interactions vs. Additive Spreads: SEADS’s admitted factors are overwhelmingly multiplicative interactions or ratios, whereas RD-Agent(Q) and Beyond Prompting mostly admit flat additive/subtractive spreads of two characteristics.The comparison concerns admitted factors pooled across SEADS’s locked-gate flagship replicates, with Table 8 providing the structural classification.
  • H Formula-Complexity Case Study: Interactions vs. Additive Spreads: Each SEADS example requires multiple ranked legs to score well simultaneously, unlike baseline spreads whose ingredients are treated as independent and linear.A flat difference is already spanned by including both characteristics separately in a linear downstream model, motivating SEADS’s triviality rejection rule.
  • H Formula-Complexity Case Study: Interactions vs. Additive Spreads: Beyond Prompting’s repeated cash-flow-yield spreads yield mean pairwise |ρ|=0.281 and an effective independent count of 3.46 among 94 nominal admits.The examples vary one cash-flow-yield ingredient against many unrelated characteristics, illustrating the redundancy that the triviality rule is designed to guard against.
  • I Broader Related-Work Landscape: Earlier factor-discovery research spans human–AI mining, reinforcement-learning search, evolutionary mining, prompted conceptualization, and semi-automated feature engineering [Wang et al., 2023; Yu et al., 2023; Cui et al., 2021; Cheng and Tang, 2024; Wang et al., 2024c].Adjacent LLM miners contribute mechanisms including chain-based generation, MCTS, GFlowNets, and on-graph evolution; the chain-based system’s reported numbers are not treated as verifiable because it was removed over a licensing-rights issue [Cao et al., 2025; Shi et al., 2026b; Chen et al., 2026a; Guo et al., 2026].
  • I Broader Related-Work Landscape: The paper distinguishes its memory tiers from skills-and-experience and memory-driven discovery precedents by the granularity and pipeline-stage specificity of feedback [Wang et al., 2026; Liu et al., 2026a].It also separates structured numerical-data factor discovery on two equity panels from text-seeded, news-based, disclosure-prediction, and debate-based systems [Lin et al., 2026; Liu et al., 2026a; Wang et al., 2024b; Choi et al., 2025; Zhao et al., 2025].
  • I Broader Related-Work Landscape: Evaluation infrastructure covers formulaic alphas and trading agents, while a surveyed median test window of 1.3 years motivates the paper’s realism argument [Ding et al., 2026; Li et al., 2025a; Ding et al., 2024].Huang et al. [2026] is identified as the closest prior instance because it uses an out-of-sample single static split with one modern backbone, but evaluates cryptocurrency rather than equities.
  • J Reproducibility Notes: Headline results use untouched out-of-sample windows, exclude those windows from every gate decision, and rely on deterministic validation, checks, fingerprinting, and sign backstops.The validation engine depends only on the candidate series, in-sample window, and threshold dictionary, preventing discovery or selection leakage.
  • J Reproducibility Notes: All systems share identical portfolio-construction and metric implementations, while SEADS and ablations report means ± standard error over independent replicates rather than single runs.The reported setup uses one current LLM backbone across roles; GPT-5.2 is called through the standard chat-completions API without tools, function calling, or web access, with role-specific temperatures (Table 9).

K Metric Definitions

This section defines the factor-evaluation metrics and gate thresholds, including Rank-IC/ICIR, coverage, regime consistency, and multiple-error controls. It also identifies tables reporting analytic and cross-replicate standard errors.

  • K Metric Definitions: Rank-IC is the monthly cross-sectional Spearman correlation between factor values and next-month excess returns, requiring at least five valid pairs and treating months with fewer than 12 correlations as undefined.ICIR is defined from the monthly Rank-IC series, whose averaging expression is truncated in the supplied passage.
  • K Metric Definitions: The locked gate applies a cumulative-candidate Bonferroni correction only to its t-statistic, with a floor of 3.0, while the Fama–MacBeth threshold remains fixed at t ≥1.0.The corrected threshold uses α = 0.05 and family-wise error-rate control through N, the cumulative candidate count.
  • K Metric Definitions: Coverage is the fraction of in-sample months with non-null candidate values meeting Min Coverage, while regime consistency recomputes mean Rank-IC on regime-specific data.The supplied passage truncates the full regime-consistency definition.
  • K Metric Definitions: Table 10 reports analytic Lo-(2002) standard errors for cSR across window lengths, including Panel B with n=59.The supplied caption identifies the table’s scope but provides no cell values.
  • K Metric Definitions: Table 11 reports per-entry cross-replicate standard errors for Table 1 across the same systems, metrics, and panels.The supplied caption provides no numerical standard-error values.

L Example Admits · M AI Acknowledgement

The section presents three hand-selected SEADS factor admits illustrating stability, low reference-panel correlation, and multiplicative structure, while acknowledging that LLMs drove every reported experiment and assisted parts of the project’s development and writing.

  • L Example Admits: Three example SEADS admits were selected to illustrate strong in-sample/out-of-sample agreement, low reference-panel correlation, and SEADS’s typical multiplicative-interaction structure.All three cleared every check in Table 6, with in-sample statistics computed at admission and out-of-sample statistics computed on the held-out window.
  • L Example Admits: Table 12 reports Coverage, Rank-IC (ICIR), Partial-IC t, FM t, and Max corr in-sample, alongside three out-of-sample-labeled rows for each example admit.The in-sample values are computed at admission time, whereas the out-of-sample values come from the held-out window.
  • L Example Admits: The stable liquidity efficiency admit demonstrates strong in-sample/out-of-sample agreement through a three-way product of trading-friction and dollar-volume-stability signals.Tight spreads and low trading inactivity indicate lower frictions and faster price discovery, while unstable dollar-volume regimes are downweighted; the admit correlates most with dolvol var 126d.
  • L Example Admits: The distress amplified perf mispricing admit has the lowest pooled correlation, 0.449 with debt at, among all admits across five replicates.Its rank(mispricing perf) × rank(o score) structure represents distress-amplified performance mispricing, where limits-to-arbitrage and forced trading may strengthen subsequent correction.
  • L Example Admits: The cash operating profitability × low tax payable growth admit exemplifies SEADS’s dominant structure, shared by 83.8% of its admits.Its two-variable rank product is rank(cop bev) × (1-rank(txp gr1a)), and it correlates most with txp gr1a.
  • M AI Acknowledgement: Language models were the experimental subject: SEADS and all five baselines used an LLM at every proposal, evaluation, and evolution step.Separately, generative AI tools assisted parts of supporting-code development and manuscript preparation.
Loading 2609.00731v1…