pith. sign in

arxiv: 2605.16895 · v1 · pith:EAKE5DYUnew · submitted 2026-05-16 · 💻 cs.CE · cs.AI· cs.CL

The Alpha Illusion: Reported Alpha from LLM Trading Agents Should Not Be Treated as Deployment Evidence

Yuxuan Ye , Jun Han , Ao Hu , Juncheng Bu , Yiyi Chen , Liangjian Wen , Danilo Mandic , Danny Dongning Sun
show 2 more authors
Xu Yinghui Zenglin Xu
This is my paper

Pith reviewed 2026-05-19 19:14 UTC · model grok-4.3

classification 💻 cs.CE cs.AIcs.CL
keywords LLM trading agentsalpha reportingdeployment validitySharpe ratiofinancial AIreporting protocolstrading benchmarks
0
0 comments X p. Extension
pith:EAKE5DYU Add to your LaTeX paper What is a Pith Number? 
\usepackage{pith}
\pithnumber{EAKE5DYU}

Prints a linked pith:EAKE5DYU badge after your title and writes the identifier into PDF metadata. Compiles on arXiv with no extra files. Learn more

The pith

Reported alpha from LLM trading agents should not be treated as deployment evidence.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper claims that headline performance numbers from end-to-end LLM trading agents, including high Sharpe ratios, do not yet prove the systems are ready for real trading use. A sympathetic reader would care because acting on these numbers without extra checks risks losses from unaccounted timing problems or execution gaps. The authors explain that existing reports mix genuine skill with issues such as data contamination over time, unmodeled costs, and narrative fitting. They introduce a set of minimum reporting protocols and a safer design that keeps LLMs as information providers rather than full trading decision makers.

Core claim

Reported alpha from end-to-end LLM trading agents should not be treated as deployment evidence. Before such returns can support claims of deployable trading capability, they must survive structural validity tests for temporal integrity, real-world frictions, counterfactual robustness, predictive calibration, numerical execution, and multi-agent disaggregation. Current public evidence cannot yet distinguish robust predictive ability from temporal contamination, unmodeled frictions, short-window Sharpe uncertainty, narrative fitting, and parametric priors. The problem is not only evaluative but structural. Language confidence is not tradable probability, narrative reasoning is not numerical n,

What carries the argument

A minimum reporting protocol suite P1-P6 with tiered applicability by claim strength, plus a modular alternative that uses LLMs as auditable information interfaces upstream of independent calibration, risk, and execution modules.

If this is right

  • Reported returns require explicit checks for temporal integrity to exclude data leakage.
  • Real-world frictions including transaction costs must be modeled before any deployment claim.
  • Counterfactual robustness tests are needed to confirm decisions hold under varied conditions.
  • Predictive calibration must ensure stated confidence matches observed outcome frequencies.
  • Multi-agent disaggregation is required to isolate individual component contributions.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Existing LLM trading benchmarks may require updates to incorporate these validity criteria.
  • The modular design offers a practical way to add language models to trading pipelines without full end-to-end reliance.
  • Standardized test harnesses implementing the proposed protocols could become a community resource for future studies.

Load-bearing premise

Current public evidence cannot yet distinguish robust predictive ability from temporal contamination, unmodeled frictions, short-window Sharpe uncertainty, narrative fitting, and parametric priors.

What would settle it

Demonstration of an LLM trading agent whose returns survive all six validity tests and then deliver positive risk-adjusted performance in a live market with real capital and frictions would settle the claim.

Figures

Figures reproduced from arXiv: 2605.16895 by Ao Hu, Danilo Mandic, Danny Dongning Sun, Juncheng Bu, Jun Han, Liangjian Wen, Xu Yinghui, Yiyi Chen, Yuxuan Ye, Zenglin Xu.

Figure 1
Figure 1. Figure 1: One-year reproduction of TradingAgents (TA) and QuantAgent (QA) vs. Buy-and-Hold [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Heatmap of 95% CI half-width of period Sharpe under Lo [15] (eq. 2, K=1), over SRc ∈ [0, 8] and T ∈ [20, 500] daily observations. The dashed line marks where the 95% CI just covers zero. Overlaid headline Sharpes from TradingAgents, FinMem, FinAgent, and FinCon [30] are illustrative (FinCon: single-stock NFLX 2.37 and best portfolio (TSLA, MSFT, PFE) 3.27); QuantAgent is omitted because it reports HFT dire… view at source ↗
Figure 3
Figure 3. Figure 3: Reported Sharpe per (system × ticker) for four LLM trading systems. Bars show headline Sharpe as published by FinMem (6 mo, Oct 2022 – Apr 2023), TradingAgents (3 mo, Jan – Mar 2024), FinAgent (7 mo, Jun 2023 – Jan 2024), and FinCon [30] (8.5 mo, Oct 2022 – Jun 2023). All numbers are gross of friction except FinAgent, which models commission only. B Friction Coverage and Per-Asset Metrics from the Reproduc… view at source ↗
Figure 4
Figure 4. Figure 4: Friction-modeling coverage of FinMem, TradingAgents, FinAgent, FinCon, and QuantA [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Two faces of parametric prior lock-in, after Lee et al. [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗
read the original abstract

End-to-end LLM trading agents have moved quickly from research curiosity to a small ecosystem of named systems, including FinCon, FinMem, TradingAgents, FinAgent, QuantAgent, and FLAG-Trader. Several of these report headline Sharpe ratios that would be material if read at face value on a deployment desk, and associated benchmarks such as FinBen report trading-task Sharpe statistics in the same range. The gap between architecture research and deployment claim has been crossed too freely on both sides of the academia--industry divide. We take a position on that gap: reported alpha from end-to-end LLM trading agents should not be treated as deployment evidence. Before such returns can support claims of deployable trading capability, they must survive structural validity tests for temporal integrity, real-world frictions, counterfactual robustness, predictive calibration, numerical execution, and multi-agent disaggregation. Current public evidence cannot yet distinguish robust predictive ability from temporal contamination, unmodeled frictions, short-window Sharpe uncertainty, narrative fitting, and parametric priors. The problem is not only evaluative but structural. Language confidence is not tradable probability, narrative reasoning is not numerical execution, and model priors may become undisclosed implicit factor exposures. We contribute a minimum reporting protocol suite, P1--P6, with tiered applicability by claim strength, and a conservative modular alternative that uses LLMs as auditable information interfaces upstream of independent calibration, risk, and execution modules. Code and reproduction harness: \url{https://github.com/hj1650782738/Trading}.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The manuscript claims that headline performance metrics (e.g., Sharpe ratios) reported by end-to-end LLM trading agents such as FinCon, FinMem, TradingAgents, FinAgent, QuantAgent, and FLAG-Trader, as well as benchmarks like FinBen, should not be treated as deployment evidence. It enumerates structural validity requirements—temporal integrity, real-world frictions, counterfactual robustness, predictive calibration, numerical execution, and multi-agent disaggregation—and states that current public evidence cannot yet distinguish robust predictive ability from temporal contamination, unmodeled frictions, short-window Sharpe uncertainty, narrative fitting, and parametric priors. The authors contribute a tiered minimum reporting protocol (P1–P6) and a conservative modular architecture that positions LLMs as auditable information interfaces upstream of independent calibration, risk, and execution modules.

Significance. If the position holds, the work supplies a timely, principle-based corrective to the rapid expansion of LLM trading-agent papers by clarifying the evidentiary threshold between architectural results and deployable alpha. The P1–P6 protocol and modular alternative constitute constructive, actionable contributions that could raise reporting standards without requiring new empirical results from the authors themselves. The argument correctly invokes standard quantitative-finance evaluation criteria rather than deriving performance from the authors’ own fitted parameters.

major comments (1)
  1. [Section discussing public evidence and the gap between architecture research and deployment claim] Discussion of cited systems (FinCon, FinMem, TradingAgents, FinAgent, QuantAgent, FLAG-Trader): the central claim that 'current public evidence cannot yet distinguish robust predictive ability from temporal contamination...' is load-bearing yet rests on the absence of explicit robustness statements rather than documented properties of the specific numbers. The manuscript would be strengthened by at least one concrete audit or citation—for example, confirming whether any reported backtest used non-overlapping temporal splits or modeled bid-ask spreads—rather than general inference from missing statements.
minor comments (2)
  1. [Abstract] Abstract: the phrasing 'the gap between architecture research and deployment claim has been crossed too freely on both sides' risks implying authorial intent; a neutral rephrasing focused solely on the evidentiary shortfall would improve precision.
  2. [Proposal of minimum reporting protocol suite] P1–P6 protocol: while the tiered applicability is a useful contribution, the manuscript would benefit from a brief illustrative application of one or two protocols to an existing cited system to demonstrate practicality.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the positive assessment and recommendation of minor revision. The single major comment is addressed point-by-point below with a commitment to strengthen the evidentiary presentation while preserving the manuscript's core argument.

read point-by-point responses

  1. Referee: Discussion of cited systems (FinCon, FinMem, TradingAgents, FinAgent, QuantAgent, FLAG-Trader): the central claim that 'current public evidence cannot yet distinguish robust predictive ability from temporal contamination...' is load-bearing yet rests on the absence of explicit robustness statements rather than documented properties of the specific numbers. The manuscript would be strengthened by at least one concrete audit or citation—for example, confirming whether any reported backtest used non-overlapping temporal splits or modeled bid-ask spreads—rather than general inference from missing statements.

    Authors: We agree that the claim benefits from greater specificity. We have re-reviewed the methodology sections of the six cited papers and confirm that none explicitly report non-overlapping temporal splits, bid-ask spread modeling, or slippage in their backtests; several rely on daily or lower-frequency data without stating execution assumptions. We will add a concise supplementary table (or expanded footnote) that maps each system to the presence or absence of these elements as stated in the original publications. This converts the inference from pure absence into a documented literature review. A full re-audit of non-public code or proprietary data remains outside the scope of this work, but the proposed addition directly addresses the request for concrete citation. revision: yes

Circularity Check

0 steps flagged

No circularity: critique relies on external quant-finance benchmarks rather than internal re-derivation

full rationale

The paper enumerates established validity tests (temporal integrity, frictions, counterfactual robustness, etc.) drawn from quantitative finance literature and proposes an independent reporting protocol P1-P6. No equations, fitted parameters, or self-citations reduce the central claim to quantities defined by the authors' own priors or prior work. The assertion that public LLM-agent results fail to distinguish skill from contamination is an inference from missing robustness statements in the cited systems, not a self-referential loop or renamed fit. The derivation chain is self-contained against external standards.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The paper draws on standard domain assumptions from quantitative finance about what counts as valid trading evidence; no new free parameters or invented entities are introduced.

axioms (2)
  • domain assumption Language confidence is not tradable probability
    Invoked in the abstract as a structural distinction between model output and executable trading signals.
  • domain assumption Narrative reasoning is not numerical execution
    Stated explicitly in the abstract as one of the core problems preventing direct use of LLM outputs for trading.

pith-pipeline@v0.9.0 · 5841 in / 1244 out tokens · 40433 ms · 2026-05-19T19:14:14.265450+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

41 extracted references · 41 canonical work pages · 4 internal anchors

  1. [1]

    Optimal execution of portfolio transactions.Journal of Risk, 3(2):5–39, 2000

    Robert Almgren and Neil Chriss. Optimal execution of portfolio transactions.Journal of Risk, 3(2):5–39, 2000

    work page 2000

  2. [2]

    StockBench: Can LLM agents trade stocks profitably in real-world markets?arXiv preprint arXiv:2510.02209, 2025

    Yanxu Chen, Zijun Yao, Yantao Liu, Amy Xin, Jin Ye, Jianing Yu, Lei Hou, and Juanzi Li. StockBench: Can LLM agents trade stocks profitably in real-world markets?arXiv preprint arXiv:2510.02209, 2025

    work page arXiv 2025

  3. [3]

    Empirical asset pricing via machine learning.The Review of Financial Studies, 33(5):2223–2273, 2020

    Shihao Gu, Bryan Kelly, and Dacheng Xiu. Empirical asset pricing via machine learning.The Review of Financial Studies, 33(5):2223–2273, 2020

    work page 2020

  4. [4]

    Weinberger

    Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. On calibration of modern neural networks. InInternational Conference on Machine Learning (ICML), pages 1321–1330. PMLR, 2017

    work page 2017

  5. [5]

    Harvey, Yan Liu, and Heqing Zhu

    Campbell R. Harvey, Yan Liu, and Heqing Zhu. . . . and the cross-section of expected returns. The Review of Financial Studies, 29(1):5–68, 2016

    work page 2016

  6. [6]

    Ojha, Ross Spoon, Jiatong Han, and Colin F

    Thomas Henning, Siddhartha M. Ojha, Ross Spoon, Jiatong Han, and Colin F. Camerer. LLM agents do not replicate human market traders: Evidence from experimental finance.arXiv preprint arXiv:2502.15800, 2025

    work page arXiv 2025

  7. [7]

    The losing winner: An LLM agent that predicts the market but loses money

    Youwon Jang, Joochan Kim, and Byoung-Tak Zhang. The losing winner: An LLM agent that predicts the market but loses money. InNeurIPS 2025 Workshop on Generative AI in Finance,

    work page 2025

  8. [8]

    OpenReview:https://openreview.net/forum?id=FzahgVWy59

    work page

  9. [9]

    Language Models (Mostly) Know What They Know

    Saurav Kadavath, Tom Conerly, Amanda Askell, Tom Henighan, Dawn Drain, Ethan Perez, Nicholas Schiefer, Zac Hatfield-Dodds, Nova DasSarma, Eli Tran-Johnson, Scott Johnston, Sheer El-Showk, Andy Jones, Nelson Elhage, Tristan Hume, Anna Chen, Yuntao Bai, Sam Bowman, Stanislav Fort, Deep Ganguli, Danny Hernandez, Josh Jacobson, Jackson Kernion, Shauna Kravec,...

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  10. [10]

    Deficiency of large language models in finance: An empirical examination of hallucination.arXiv preprint arXiv:2311.15548, 2023

    Haoqiang Kang and Xiao-Yang Liu. Deficiency of large language models in finance: An empirical examination of hallucination.arXiv preprint arXiv:2311.15548, 2023

    work page arXiv 2023

  11. [11]

    Your AI, not your view: The bias of LLMs in investment analysis.arXiv preprint arXiv:2507.20957, 2025

    Hoyoung Lee, Junhyuk Seo, Suhwan Park, Junhyeong Lee, Wonbin Ahn, Chanyeol Choi, Alejandro Lopez-Lira, and Yongjae Lee. Your AI, not your view: The bias of LLMs in investment analysis.arXiv preprint arXiv:2507.20957, 2025. ACM International Conference on AI in Finance (ICAIF) 2025

    work page arXiv 2025

  12. [12]

    Profit mirage: Revisiting information leakage in LLM-based financial agents.arXiv preprint arXiv:2510.07920, 2025

    Xiangyu Li, Yawen Zeng, Xiaofen Xing, Jin Xu, and Xiangmin Xu. Profit mirage: Revisiting information leakage in LLM-based financial agents.arXiv preprint arXiv:2510.07920, 2025

    work page arXiv 2025

  13. [13]

    Troubling Trends in Machine Learning Scholarship

    Zachary C. Lipton and Jacob Steinhardt. Troubling trends in machine learning scholarship. arXiv preprint arXiv:1807.03341, 2018. Presented at ICML 2018: The Debates

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  14. [14]

    AgentBench: Evaluating LLMs as Agents

    Xiao Liu, Hao Yu, Hanchen Zhang, Yifan Xu, Xuanyu Lei, Hanyu Lai, Yu Gu, Hangliang Ding, Kaiwen Men, Kejuan Yang, Shudan Zhang, Xiang Deng, Aohan Zeng, Zhengxiao Du, Chenhui Zhang, Sheng Shen, Tianjun Zhang, Yu Su, Huan Sun, Minlie Huang, Yuxiao Dong, and Jie Tang. AgentBench: Evaluating LLMs as agents. InInternational Conference on Learning Representatio...

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  15. [15]

    arXiv preprint arXiv:2112.06753 , year =

    Xiao-Yang Liu, Jingyang Rui, Jiechao Gao, Liuqing Yang, Hongyang Yang, Zhaoran Wang, Christina Dan Wang, and Jian Guo. FinRL-Meta: A universe of near-real market environments for data-driven deep reinforcement learning in quantitative finance. InNeurIPS 2021 Workshop on Data-Centric AI, 2021. arXiv:2112.06753

    work page arXiv 2021

  16. [16]

    Andrew W. Lo. The statistics of Sharpe ratios.Financial Analysts Journal, 58(4):36–52, 2002. 10

    work page 2002

  17. [17]

    Can ChatGPT forecast stock price movements? Return predictability and large language models.arXiv preprint arXiv:2304.07619, 2023

    Alejandro Lopez-Lira and Yuehua Tang. Can ChatGPT forecast stock price movements? Return predictability and large language models.arXiv preprint arXiv:2304.07619, 2023

    work page arXiv 2023

  18. [18]

    FinToolBench: Evaluating LLM agents for real-world financial tool use.arXiv preprint arXiv:2603.08262, 2026

    Jiaxuan Lu, Kong Wang, Yemin Wang, Qingmei Tang, Hongwei Zeng, Xiang Chen, Jiahao Pi, Shujian Deng, Lingzhi Chen, Yi Fu, Kehua Yang, and Xiao Sun. FinToolBench: Evaluating LLM agents for real-world financial tool use.arXiv preprint arXiv:2603.08262, 2026

    work page arXiv 2026

  19. [19]

    Agent trading arena: A study on numerical understanding in LLM-based agents.arXiv preprint arXiv:2502.17967, 2025

    Tianmi Ma, Jiawei Du, Wenxin Huang, Wenjie Wang, Liang Xie, Xian Zhong, and Joey Tianyi Zhou. Agent trading arena: A study on numerical understanding in LLM-based agents.arXiv preprint arXiv:2502.17967, 2025

    work page arXiv 2025

  20. [20]

    A fast and effective solution to the problem of look-ahead bias in LLMs.arXiv preprint arXiv:2512.06607, 2025

    Humzah Merchant and Bradford Levy. A fast and effective solution to the problem of look-ahead bias in LLMs.arXiv preprint arXiv:2512.06607, 2025

    work page arXiv 2025

  21. [21]

    Robert Novy-Marx and Mihail Z. Velikov. AI-powered (finance) scholarship. NBER Working Paper 33363, National Bureau of Economic Research, 2025

    work page 2025

  22. [22]

    Obizhaeva and Jiang Wang

    Anna A. Obizhaeva and Jiang Wang. Optimal trading strategy and supply/demand dynamics. Journal of Financial Markets, 16(1):1–32, 2013

    work page 2013

  23. [23]

    When agents trade: Live multi-market trading benchmark for LLM agents.arXiv preprint arXiv:2510.11695, 2025

    Lingfei Qian, Xueqing Peng, Yan Wang, Vincent Jim Zhang, Huan He, Hanley Smith, Yi Han, Yueru He, Haohang Li, Yupeng Cao, Yangyang Yu, Alejandro Lopez-Lira, Peng Lu, Jian-Yun Nie, Guojun Xiong, Jimin Huang, and Sophia Ananiadou. When agents trade: Live multi-market trading benchmark for LLM agents.arXiv preprint arXiv:2510.11695, 2025

    work page arXiv 2025

  24. [24]

    Beyond the reported cutoff: Where large language models fall short on financial knowledge

    Agam Shah, Liqin Ye, Sebastian Jaskowski, Wei Xu, and Sudheer Chava. Beyond the reported cutoff: Where large language models fall short on financial knowledge. InConference on Language Modeling (COLM), 2025. arXiv:2504.00042

    work page arXiv 2025

  25. [25]

    Algorithms with calibrated ma- chine learning predictions

    Judy Hanwen Shen, Ellen Vitercik, and Anders Wikum. Algorithms with calibrated ma- chine learning predictions. InInternational Conference on Machine Learning (ICML), 2025. arXiv:2502.02861

    work page arXiv 2025

  26. [26]

    TradingAgents: Multi-agents LLM financial trading framework.arXiv preprint arXiv:2412.20138, 2024

    Yijia Xiao, Edward Sun, Di Luo, and Wei Wang. TradingAgents: Multi-agents LLM financial trading framework.arXiv preprint arXiv:2412.20138, 2024

    work page arXiv 2024

  27. [27]

    FinBen: A holistic financial benchmark for large language models

    Qianqian Xie, Weiguang Han, Zhengyu Chen, Ruoyu Xiang, Xiao Zhang, Yueru He, Mengxi Xiao, Dong Li, Yongfu Dai, Duanyu Feng, Yijing Xu, Haoqiang Kang, Ziyan Kuang, Chenhan Yuan, Kailai Yang, Zheheng Luo, Tianlin Zhang, Zhiwei Liu, Guojun Xiong, Zhiyang Deng, Yuechen Jiang, Zhiyuan Yao, Haohang Li, Yangyang Yu, Gang Hu, Jiajia Huang, Xiao-Yang Liu, Alejandr...

    work page arXiv 2024

  28. [28]

    QuantAgent: Price-driven multi-agent LLMs for high-frequency trading.arXiv preprint arXiv:2509.09995, 2025

    Fei Xiong, Xiang Zhang, Aosong Feng, Siqi Sun, and Chenyu You. QuantAgent: Price-driven multi-agent LLMs for high-frequency trading.arXiv preprint arXiv:2509.09995, 2025

    work page arXiv 2025

  29. [29]

    Smith, Xiao-Yang Liu, Jimin Huang, Sophia Ananiadou, and Qianqian Xie

    Guojun Xiong, Zhiyang Deng, Keyi Wang, Yupeng Cao, Haohang Li, Yangyang Yu, Xueqing Peng, Mingquan Lin, Kaleb E. Smith, Xiao-Yang Liu, Jimin Huang, Sophia Ananiadou, and Qianqian Xie. FLAG-Trader: Fusion LLM-agent with gradient-based reinforcement learning for financial trading.arXiv preprint arXiv:2502.11433, 2025

    work page arXiv 2025

  30. [30]

    Suchow, and Khaldoun Khashanah

    Yangyang Yu, Haohang Li, Zhi Chen, Yuechen Jiang, Yang Li, Denghui Zhang, Rong Liu, Jordan W. Suchow, and Khaldoun Khashanah. FinMem: A performance-enhanced LLM trading agent with layered memory and character design.arXiv preprint arXiv:2311.13743, 2023

    work page arXiv 2023

  31. [31]

    Suchow, Rong Liu, Zhenyu Cui, Zhaozhuo Xu, Denghui Zhang, Koduvayur Subbalakshmi, Guojun Xiong, Yueru He, Jimin Huang, Dong Li, and Qianqian Xie

    Yangyang Yu, Zhiyuan Yao, Haohang Li, Zhiyang Deng, Yupeng Cao, Zhi Chen, Jordan W. Suchow, Rong Liu, Zhenyu Cui, Zhaozhuo Xu, Denghui Zhang, Koduvayur Subbalakshmi, Guojun Xiong, Yueru He, Jimin Huang, Dong Li, and Qianqian Xie. FinCon: A synthe- sized LLM multi-agent system with conceptual verbal reinforcement for enhanced financial decision making. InA...

    work page arXiv 2024

  32. [32]

    TraderBench: How robust are AI agents in adversarial capital markets?arXiv preprint arXiv:2603.00285, 2026

    Xiaochuang Yuan, Hui Xu, Silvia Xu, Cui Zou, and Jing Xiong. TraderBench: How robust are AI agents in adversarial capital markets?arXiv preprint arXiv:2603.00285, 2026

    work page arXiv 2026

  33. [33]

    Stop overvaluing multi-agent debate—we must rethink evaluation and embrace model heterogeneity.arXiv preprint arXiv:2502.08788, 2025

    Hangfan Zhang, Zhiyao Cui, Jianhao Chen, Xinrun Wang, Qiaosheng Zhang, Zhen Wang, Dinghao Wu, and Shuyue Hu. Stop overvaluing multi-agent debate—we must rethink evaluation and embrace model heterogeneity.arXiv preprint arXiv:2502.08788, 2025

    work page arXiv 2025

  34. [34]

    A multimodal foundation agent for financial trading: Tool-augmented, diversified, and generalist

    Wentao Zhang, Lingxuan Zhao, Haochong Xia, Shuo Sun, Jiaze Sun, Molei Qin, Xinyi Li, Yuqing Zhao, Yilei Zhao, Xinyu Cai, Longtao Zheng, Xinrun Wang, and Bo An. A multimodal foundation agent for financial trading: Tool-augmented, diversified, and generalist. InProceed- ings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD),

    work page

  35. [35]

    WebArena: A Realistic Web Environment for Building Autonomous Agents

    Shuyan Zhou, Frank F. Xu, Hao Zhu, Xuhui Zhou, Robert Lo, Abishek Sridhar, Xianyi Cheng, Tianyue Ou, Yonatan Bisk, Daniel Fried, Uri Alon, and Graham Neubig. WebArena: A realistic web environment for building autonomous agents. InInternational Conference on Learning Representations (ICLR), 2024. arXiv:2307.13854. A Per-Ticker Reported Sharpe across Three ...

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  36. [36]

    Information ex- traction LLM-led, schema-bound Extract structured propositions from news, filings, calls, announce- ments Source span, publication time, re- trieval time, entity/event labels

    work page

  37. [37]

    Feature construc- tion Quant module Provides candidate structured in- puts, does not set feature weights Point-in-time feature table, missing- data handling, normalization rules

    work page

  38. [38]

    Signal synthesis Quant model One of many information sources Ablation and marginal contribution of LLM-derived features

    work page

  39. [39]

    Probability cali- bration Independent statistical module Does not provide self-rated trading probabilities ECE, reliability curves, regime- conditioned calibration, out-of- sample stability

    work page

  40. [40]

    Sizing and risk control Portfolio and risk mod- ules Explains sources of risk, does not directly determine size Factor exposures, sector exposures, leverage, drawdown, liquidity con- straints

    work page

  41. [41]

    Suppose at time t an 8-K is released stating that company X is cutting next-quarter revenue guidance

    Execution and au- dit Execution system Observer or summarizer Order log, fill prices, slippage, latency, failed orders, risk-stop records E Worked Example: 8-K Guidance Cut Through the Modular Pipeline This appendix walks the modular stage structure (Table 5) on a concrete event, as a description of existing practice rather than a proposal. Suppose at tim...

    work page