ContestTrade: A Multi-Agent Trading System Based on Internal Contest Mechanism

Bo Yang; Jing Li; Li Zhao; Mengting Chen; Rui Sun; Xinyang Wang; Yuxiao Bai; Zuo Bai; Zuoyou Jiang

REVIEW 4 major objections 5 minor 33 references

ContestTrade: A Multi-Agent Trading System Based on Internal Contest Mechanism

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read ContestTrade uses internal agent competition to beat multi-agent and quant baselines on A-share trading, reporting 52.80% cumulative return and a 3.12 Sharpe ratio.

desk verdict A plausible contest-based multi-agent trading architecture whose headline results are likely lookahead artifacts due to an unspecified score-timing lag; the idea is worth reviewing but the evidence needs a rework. read the letter →

arxiv 2508.00554 v3 pith:R5W76BSN submitted 2025-08-01 q-fin.TR cs.CLq-fin.CP

Li Zhao , Rui Sun , Zuoyou Jiang , Bo Yang , Yuxiao Bai , Mengting Chen , Xinyang Wang , Jing Li

show 1 more author

Zuo Bai

This is my paper

classification q-fin.TRcs.CLq-fin.CP

keywords internalcontestmechanismmulti-agenttradingLLMagentstextualfactorsA-sharemarketportfolioconstructionnoisezero-intelligencetrader

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper proposes ContestTrade, a two-team LLM-based trading system in which Data Analysts and Research Agents compete internally, and only the top performers' outputs are used. It claims this internal contest mechanism materially improves robustness to market noise compared with fixed multi-agent systems and traditional quant methods. On daily A-share trading from January to June 2025, it reports a cumulative return of 52.80%, a Sharpe ratio of 3.12, and a maximum drawdown of 12.41%, alongside Rank IC values of 0.054 and 0.079 for the two contests. If these numbers hold up, the contest mechanism is a practical way to turn noisy LLM output into a stable trading signal.

What carries the argument

The load-bearing object is the internal contest mechanism, a three-phase 'Quantify-Predict-Allocate' pipeline that maps a set of agents to a capital or factor weight vector. Quantify turns each agent's recent output into a number: for Data Agents, a zero-intelligence trader sums rating times next-day price change over the factor's observations; for Research Agents, a hybrid of trailing Sharpe and LLM-judge quality grades. Predict exploits short-term momentum in these scores, using a simple gradient-boosting model with a five-day lookback to forecast three-day (factor) or five-day (strategy) utility. Allocate then enforces the context constraint: the data portfolio is solved as a 0/1 knapsack under a 16k-token budget, while research capital is spread proportionally to positive predicted Sharpe ratios.

What would settle it

Run the pipeline in a strict real-time simulation where no price from day t+1 can enter any score computed on day t, then compare the cumulative return and Sharpe to the reported 52.80% and 3.12. A simpler decisive check: for a sample of days, list the factors the Data Analyst Contest selected on day t and the PriceChange(t, t+1) values used to score them; if the selection is systematically correlated with the next-day returns it was scored on, the reported backtest is unachievable.

Watch

Extended reading notes

Core claim

The central discovery is that a competitive 'Quantify-Predict-Allocate' loop can be used to select, in real time, which agents' outputs should drive a trade. Each Data Agent produces a textual factor; a zero-intelligence trader scores every atomic observation by how well the named stock moved the next day, and a gradient-boosting model predicts which factors will keep working over the next few days. The factor set is then chosen as the 0/1 knapsack solution that maximises predicted risk-adjusted score under a 16,000-token context budget. Research Agents then submit structured buy/hold/sell signals, a hybrid score combining trailing performance and an LLM judge's assessment is predicted forward five days, and capital is allocated in proportion to positive predicted Sharpe ratios. The authors argue this two-stage competition is what lets the system adapt to style rotation and reject market noise.

Load-bearing premise

The entire result depends on the contest scores being computable before the trades they guide are placed; Algorithm 1 scores each observation by multiplying a rating by the next day's price change, and the paper never specifies the lag between that scoring, the agents' predictions, and actual execution.

Editorial extensions

If this is right

If the central claim is right, running the contest with more agents does not require a new decision rule; the allocation layer merely solves a larger knapsack (data) or re-weights the positive-utility agents (research).
The reported optimal short momentum windows (5-day lookback, 3-day factor horizon, 5-day strategy horizon) imply the system can rotate factor and strategy exposures as quickly as style shifts, rather than committing to a static portfolio.
Because the entire pipeline is framed around a context budget, the same mechanism transfers to any LLM whose reasoning quality decays with context length; the only calibration needed is the decay curve.
The ablation results, where removing either contest or the deep-research tool use drops cumulative return from 52.80% to between 3.01% and 43.75%, indicate the contest selection itself is a substantial source of the reported performance.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A test the paper leaves implicit: the same Quantify-Predict-Allocate design should be applied to non-trading LLM workflows with delayed, objectively scoreable outcomes, such as automated research or content curation, to see whether the contest, not the market, is the source of the gain.
The reported six-month window is a single bull-market regime; without a downturn or regime change, it is unknown whether the contest mechanism's adaptivity helps in stress periods, which is precisely when the paper claims noise filtering matters most.
The lookahead risk in Algorithm 1 is the first thing an independent replication should check: if the scoring reward uses next-day returns and the trade is placed the same day, the reported returns are not achievable; the paper does not state the scoring-to-execution lag.
A stronger check of the Rank IC values would be to compare the daily IC distribution against zero with standard errors; values of 0.054 and 0.079 are small enough that their statistical significance over only six months should be quantified.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Take a look at this one if you track LLM trading agents, but don't take the headline numbers at face value. The real story is the timing ambiguity in Algorithm 1. The ZI trader score for an observation uses PriceChange(s.code, t+1), and Eq. (5) sets q_{i,t} to the sum of such rewards. As written, q_{i,t} can only be computed after the t+1 return is known. Allocation in Eq. (7) is defined at time t and the portfolio is rebuilt every n days. If the score used to allocate at t is q_{i,t}, then factor selection uses the first day's return of the very holding period it is trying to predict. That makes the reported 52.80% cumulative return a lookahead artifact. The paper never states the lag between scoring, prediction, and execution. This is the load-bearing issue.

What is actually new: the internal contest idea—score agents with a zero-intelligence trader, predict future utility with LightGBM, then allocate to top factors/signals—is applied inside a two-team LLM pipeline. That is a distinct mechanism, not just another ensemble. Framing factor selection as a knapsack problem under an effective context constraint is also sensible, and the ablations are the right ones: removing each component degrades results in a consistent way. Credit where due: the test period is after the LLM knowledge cutoff, T+1 settlement and price limits are respected, transaction costs are included (0.001), and the code is supposedly open-sourced.

The soft spots beyond the timing: one six-month A-share period, no statistical significance tests, no sensitivity analysis on the many hyperparameters (m, n, L0, sigmoid k, transaction cost). The Rank IC/ICIR values they call remarkable are modest (ICIR 0.13 and 0.18). The baseline MASS result of -19.12% is suspicious and suggests the baselines were not carefully tuned. None of these are fatal by themselves, but combined with the lookahead issue the paper's central claim is not established.

The paper is for researchers in LLM-based trading and factor mining who want new selection mechanisms. It deserves a serious referee: the idea is interesting and the lookahead issue is fixable in principle. I would send it to review but with a strong request to clarify the timing and re-run the backtest with a properly lagged score. As is, it is a reject.

Referee Report

4 major / 5 minor

Summary. The paper proposes ContestTrade, a two-team LLM multi-agent trading system. The Data Team converts raw A-share market data into textual factors, each of which is scored by a zero-intelligence trader; the Research Team produces trading signals using deep-research tools, and each signal is scored by a combination of realized performance and an LLM judge. A "Quantify-Predict-Allocate" contest mechanism trains LightGBM models on score histories to predict future utility, then allocates factor context or capital accordingly. The authors report a 52.80% cumulative return, a 3.12 Sharpe ratio, and a 12.41% maximum drawdown on daily A-share trading from January to June 2025, with ablations suggesting that every component contributes to performance.

Significance. The architecture is plausible, and the temporal split (training July-December 2024, testing January-June 2025), the use of an open-source LLM backbone, and the open-sourced code are genuine strengths. If the timing ambiguity were resolved and the result held, the internal contest mechanism would be a meaningful contribution to LLM-based trading. However, the paper's central claim is not supported as written: Algorithm 1 may introduce lookahead, there is no statistical or multi-regime evidence for the headline numbers, and the ablation study is based on single random draws. These issues are load-bearing for the claim that ContestTrade significantly outperforms existing methods.

major comments (4)

[Contest Mechanism / Algorithm 1 / Eqs. (5)-(7)] As written, the score q_{i,t} for a factor at time t is computed from PriceChange(s.code, t+1), so q_{i,t} is not observable until after the t+1 market close. Equation (7) then uses the score history q_{i,t-m+1:t} to select the factor portfolio at time t, and the portfolio is reconstructed every n=3 days. This means the selection can use the return of the first day of the very holding period it is supposed to predict. The manuscript never specifies the lag between the factor date, the scoring date, and the allocation/execution date; the statement that model training and calibration are confined to July-December 2024 addresses parameter fitting, not real-time availability of the score. The reported 52.80% cumulative return, 3.12 Sharpe ratio, and the Rank IC values in Table 3 may therefore be lookahead artifacts. The authors should specify the exact chronological pipeline and verify with the open-sourced code that every score used in a date-t allocation is computable from data available before the allocation is made.
[Experiments / Main Results / Table 2 and Figure 4] The headline comparison rests on a single six-month test window (January-June 2025) in one market regime, and the baseline rows are single trajectories. The phrase "significantly outperforms" is not supported: no confidence intervals, standard errors, bootstrap resamples, or multiple test periods are reported. The Rank IC values in Table 3 (0.054 and 0.079, with ICIRs of 0.13 and 0.18) are presented without confidence intervals or benchmark comparisons, so calling them "strong" is not justified. The authors should add distributional evidence, such as block bootstrap over the test period or evaluation over additional out-of-sample windows, and report turnover and net-of-cost detail.
[Ablation Studies / Table 4] The variants labeled "w/o LLM Judge", "w/o Contest - Researcher", and "w/o Contest - Data Analyst" replace the contested selection with a "randomly chosen" output. Since no seeds or repeated runs are reported, the observed differences (for example, 50.55% versus 52.80% cumulative return) could be random draws. Without repetitions and error bars, the conclusion that "every component is crucial" is not established. The ablation table should report means and standard deviations over multiple random draws.
[Contest Mechanism / Eq. (6) and parameter settings] The momentum windows m=5 and n=3 for factors (and n=5 for researchers), the context lengths L0=32k and L*=16k, the sigmoid steepness k in Eq. (3), and the 0.001 transaction-cost assumption are free parameters selected on a single training period, and no sensitivity analysis is given. The claim that short-term momentum is optimal, as expressed in Eq. (6), is presented without statistical tests. Since these choices directly drive the allocations in Eqs. (7) and (8), their stability should be demonstrated before the headline result can be accepted.

minor comments (5)

[Eq. (3) and references] The citation for Eq. (3) appears as "(?Modarressi et al. 2025)" and the text also cites "Zhou et al. 2025" for the sigmoid context-capability approximation; the intended references should be corrected and formatted consistently.
[Eq. (6)] Equation (6) uses notation such as RIC(q_{t-m:t}, q_{t:t+n}) without defining the interval or the estimator for the Rank Information Coefficient; please define both.
[Researcher Contest Allocation / Eq. (8)] In Eq. (8), if all predicted Sharpe ratios are non-positive, the denominator is zero; the paper should state what the system does in that case, for example holding cash or the benchmark index.
[Experiment Setup] The experimental description does not state the number of Data Analysis Agents and Research Agents, the number of factors produced each day, or the stock universe considered; these details are necessary for reproducibility.
[Experiment Setup / Trading simulation] The paper states that the simulation adheres to T+1 settlement and daily price limits but does not describe execution assumptions, position sizing, or how "sell" signals are implemented in an A-share market context where short selling is restricted; these should be clarified.

Circularity Check

1 steps flagged · score 6.0 of 10

Data Analyst Contest's q_i,t is defined via PriceChange(t+1), while Eq. (7) selects the portfolio at t from q_i,t, so the headline backtest selection embeds the next-day return.

fitted input called prediction [Contest Mechanism > Data Analyst Contest: Algorithm 1 (ZI Trader), Eq. (5), Prediction, Eq. (7)]
"reward = s.rating × PriceChange(s.code, t + 1) ... qi,t = X obs∈Fi,t ZI(obs) (5) ... Wt = argmax Wt∈{0,1}N PN i=1 ûi,t+n · wi,t s.t. PN i=1 li · wi,t ≤ L∗ (7) ... the portfolio is reconstructed every n days"

Eq. (5) defines the date-t factor score as the sum of ZI rewards over observations in factor i 'at time t', and Algorithm 1 computes each ZI reward from PriceChange(s.code, t+1). Eq. (7) constructs the date-t factor portfolio by maximizing predicted utilities computed from the recent score sequence, x_i,t = Φ(q_i,t−m+1:t). As written, therefore, the date-t selection score for a factor contains the next day's price change. The portfolio is reconstructed every n=3 days, so the first holding-period return is already inside the score used to select that portfolio.

full rationale

ContestTrade's central proposal is a two-stage contest in which Data Analysis Agents are scored and then selected by a LightGBM prediction of future utility. The LightGBM itself is trained on July-December 2024 and evaluated on January-June 2025, which is a genuine out-of-sample split, and m,n are chosen on the training window; those parts are not circular. The paper also does not lean on self-citations: the sigmoid context-capacity model (Modarressi et al. 2025; Zhou et al. 2025), the Zero-Intelligence trader concept (Gode and Sunder 1993), and the LLM backbones are external sources. The circularity concern is concentrated in the as-written temporal semantics of the Data Analyst Contest: Algorithm 1's reward uses PriceChange(s.code, t+1), Eq. (5) aggregates that into q_i,t, and Eq. (7) uses predictions built on q_i,t to allocate at time t, with no stated lag between observation date, scoring date, and allocation date. Under the equations as written, W_t is a function of r_{t+1}, and the first day of the holding period is part of the selection signal; the reported performance is therefore, at least partially, a selection-on-realized-return artifact rather than an independent prediction. Because this affects the headline profitability claim, the circularity score is 6. If the authors intended a one-day or longer lag, the paper should specify it in Algorithm 1 and Eq. (7); as published, the claimed backtest is not causally self-contained.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central claim rests chiefly on domain assumptions about using next-day returns as ground truth and extrapolating short-term momentum. Several hyperparameters are hand-set or selected on the training window, which is legitimate but narrows the evidence. No invented physical entities are introduced.

free parameters (7)

Data Analyst Contest momentum window m = 5 days, selected on the July-December 2024 training window.
Reported as optimal from RIC analysis; not derived from theory.
Data Analyst Contest prediction horizon n = 3 days.
Chosen from RIC on the training window; portfolio is rebuilt every n days.
Researcher Contest prediction horizon n = 5 days.
Reported optimal for strategies; selected on the training window.
Effective context capacity L* = 16k tokens.
Hand-set below L0=32k to reserve context for reasoning; no ablation reported.
Context inflection point L0 = 32k tokens.
Adopted from cited prior work; not re-derived here.
Sigmoid steepness k in Eq. (3) = Not given in the paper.
Equation (3) depends on k, but no value or source is supplied; the paper proceeds directly to L*.
Transaction cost assumption = 0.001 per trade.
Fixed in all simulations; no sensitivity analysis is reported.

assumptions (5)

domain assumption Next-day price change is a valid label for the information value of a textual observation.
Algorithm 1 defines ZI reward as rating times PriceChange(s.code, t+1); this assumes market reaction measures factor quality.
domain assumption Factor and agent performance show short-term momentum that can be extrapolated by LightGBM.
The prediction section hypothesizes momentum and validates it with RIC on the training window; the extrapolation is an empirical bet.
domain assumption LLM decision capability decays sigmoidally with context length, with inflection near 32k tokens.
Equation (3) is imported from cited work; k is unspecified and L* is hand-set.
domain assumption The A-share simulation with T+1 settlement, daily price limits, and 0.001 cost is a faithful live-trading proxy.
The setup lists these rules but gives no validation of fills, slippage, or survivorship effects.
standard math The 0/1 knapsack dynamic program solves Eq. (7) exactly.
Standard algorithm; not a source of risk.

how reviews work

0 comments

Cite this review

Pith. "Pith review of ContestTrade: A Multi-Agent Trading System Based on Internal Contest Mechanism." pith.science (2026). https://pith.science/paper/R5W76BSN

@misc{pith2026250800554,
  author       = {Pith},
  title        = {Pith review of: ContestTrade: A Multi-Agent Trading System Based on Internal Contest Mechanism},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R5W76BSN}},
  note         = {Machine review of arXiv:2508.00554}
}

read the original abstract

In financial trading, large language model (LLM)-based agents demonstrate significant potential. However, the high sensitivity to market noise undermines the performance of LLM-based trading systems. To address this limitation, we propose a novel multi-agent system featuring an internal competitive mechanism inspired by modern corporate management structures. The system consists of two specialized teams: (1) Data Team - responsible for processing and condensing massive market data into diversified text factors, ensuring they fit the model's constrained context. (2) Research Team - tasked with making parallelized multipath trading decisions based on deep research methods. The core innovation lies in implementing a real-time evaluation and ranking mechanism within each team, driven by authentic market feedback. Each agent's performance undergoes continuous scoring and ranking, with only outputs from top-performing agents being adopted. The design enables the system to adaptively adjust to dynamic environment, enhances robustness against market noise and ultimately delivers superior trading performance. Experimental results demonstrate that our proposed system significantly outperforms prevailing multi-agent systems and traditional quantitative investment methods across diverse evaluation metrics. ContestTrade is open-sourced on GitHub at https://github.com/FinStep-AI/ContestTrade.

Figures

Figures reproduced from arXiv: 2508.00554 by the authors.

**Figure 2.** The Data Team Architecture, showing the work [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 4.** Portfolio value over time. This figure compares [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗

Figures from the paper (1 more)

**Figure 5.** Figure 5: Portfolio value over time. This figure compares [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 16 canonical work pages

[1]

, " * write output.state after.block = add.period write newline

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...
[2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...
[3]

Bai, J.; Bai, S.; Chu, Y.; Cui, Z.; Dang, K.; Deng, X.; Fan, Y.; Ge, W.; Han, Y.; Huang, F.; Hui, B.; et al. 2023. QWEN TECHNICAL REPORT. arXiv preprint arXiv:2309.16609

arXiv 2023
[4]

Byrd, D.; Hybinette, M.; and Balch, T. H. 2020. ABIDES: Towards High-Fidelity Multi-Agent Market Simulation. In Proceedings of the 2020 ACM SIGSIM Conference on Principles of Advanced Discrete Simulation, SIGSIM-PADS '20, 11–22. New York, NY, USA: Association for Computing Machinery. ISBN 9781450375924

work page 2020
[5]

Chroma. 2024. Context Rot: How LLMs Degrade with Context. https://research.trychroma.com/context-rot

work page 2024
[6]

DeepSeek-AI; Liu, A.; Feng, B.; Xue, B.; Wang, B.; Wu, B.; Lu, C.; Zhao, C.; Deng, C.; Zhang, C.; Ruan, C.; Dai, D.; Guo, D.; Yang, D.; Chen, D.; Ji, D.; Li, E.; Lin, F.; Dai, F.; Luo, F.; Hao, G.; Chen, G.; Li, G.; Zhang, H.; Bao, H.; Xu, H.; Wang, H.; Zhang, H.; Ding, H.; Xin, H.; Gao, H.; Li, H.; Qu, H.; Cai, J. L.; Liang, J.; Guo, J.; Ni, J.; Li, J.; ...

arXiv 2024
[7]

Ding, H.; Wang, J.; Li, Y.; and Chen, H. 2024. Large Language Model Agent in Financial Trading: A Survey. arXiv preprint arXiv:2408.06361

arXiv 2024
[8]

Engle, R. F. 1982. Autoregressive Conditional Heteroscedasticity with Estimates of the Variance of UK Inflation. Econometrica, 50(4): 987--1008

work page 1982

Show all 33 references

[9]

Fatouros, G.; Metaxas, K.; Soldatos, J.; and Kyriazis, D. 2024. Can Large Language Models Beat Wall Street? Unveiling the Potential of AI in Stock Selection . arXiv preprint arXiv:2401.03737

2024 arXiv
[10]

K.; and Sunder, S

Gode, D. K.; and Sunder, S. 1993. Allocative Efficiency of Markets with Zero-Intelligence Traders: Market as a Partial Substitute for Individual Rationality. Journal of Political Economy, 101(1): 119--137

1993
[11]

Guo, T.; Shen, H.; Huang, J.; Mao, Z.; Luo, J.; Chen, Z.; Liu, X.; Xia, B.; Liu, L.; Ma, Y.; and Zhang, M. 2025. MASS: Multi-Agent Simulation Scaling for Portfolio Construction. arXiv preprint arXiv:2505.10278

2025
[12]

J.; Ma, Y.; Ng, R.; and Chua, T.-S

Koa, K. J.; Ma, Y.; Ng, R.; and Chua, T.-S. 2024. Learning to Generate Explainable Stock Predictions using Self-Reflective Large Language Models. In The Web Conference 2024

2024
[13]

LeBaron, B. 2006. Agent-Based Computational Finance. In Tesfatsion, L.; and Judd, K. L., eds., Handbook of Computational Economics, volume 2, 1187--1233. Elsevier

2006
[14]

W.; and Khashanah, K

Li, H.; Yu, Y.; Chen, Z.; Jiang, Y.; Li, Y.; Zhang, D.; Liu, R.; Suchow, J. W.; and Khashanah, K. 2024. FinMem: A Performance-Enhanced LLM Trading Agent with Layered Memory and Character Design. In ICLR 2024 Workshop on Large Language Model (LLM) Agents

2024
[15]

Li, Y.; Yu, Y.; Li, H.; Chen, Z.; and Khashanah, K. 2023. TradingGPT: Multi-Agent System with Layered Memory and Distinct Characters for Enhanced Financial Trading Performance. Papers 2309.03736, arXiv.org

2023 arXiv
[16]

Liu, Z.; Huang, D.; Huang, K.; Li, Z.; and Zhao, J. 2021. FinBERT: a pre-trained financial language representation model for financial text mining. In Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, IJCAI'20. ISBN 9780999241165

2021
[17]

Lopez-Lira, A.; and Tang, Y. 2024. Can ChatGPT Forecast Stock Price Movements? Return Predictability and Large Language Models. arXiv preprint arXiv:2304.07619

2024
[18]

Luukkonen, R.; Komulainen, V.; Luoma, J.; Eskelinen, A.; Kanerva, J.; Kupari, H.-M.; Ginter, F.; Laippala, V.; Muennighoff, N.; Piktus, A.; Wang, T.; Tazi, N.; Scao, T.; Wolf, T.; Suominen, O.; Sairanen, S.; Merioksa, M.; Heinonen, J.; Vahtola, A.; Antao, S.; and Pyysalo, S. 2...

2023
[19]

Malkiel, B. G. 1973. A Random Walk Down Wall Street. W.W. Norton & Company. ISBN 978-0393358384

1973
[20]

A.; Yoon, S.; and Schuetze, H

Modarressi, A.; Deilamsalehy, H.; Dernoncourt, F.; Bui, T.; Rossi, R. A.; Yoon, S.; and Schuetze, H. 2025. NoLiMa: Long-Context Evaluation Beyond Literal Matching. In Forty-second International Conference on Machine Learning

2025
[21]

Wang, M.; Izumi, K.; and Sakaji, H. 2024. LLMF actor: Extracting Profitable Factors through Prompts for Explainable Stock Movement Prediction. In Ku, L.-W.; Martins, A.; and Srikumar, V., eds., Findings of the Association for Computational Linguistics: ACL 2024, 3120--3131. Ba...

2024
[22]

Wang, S.; Yuan, H.; and Ni, J., Lionel M. and Guo. 2024. QuantAgent: Seeking Holy Grail in Trading by Self-Improving Large Language Model. arXiv preprint arXiv:2402.03755

2024 arXiv
[23]

and Shum; and Guo, J

Wang, S.; Yuan, H.; Zhou, L.; Ni, H.-Y., Lionel M. and Shum; and Guo, J. 2023. Alpha-GPT: Human-AI Interactive Alpha Mining for Quantitative Investment. arXiv preprint arXiv:2308.00016

2023
[24]

Wu, S.; Irsoy, O.; Lu, S.; Dabravolski, V.; Dredze, M.; Gehrmann, S.; Kambadur, P.; Rosenberg, D.; and Mann, G. 2023. BloombergGPT: A Large Language Model for Finance. arXiv preprint arXiv:2303.17564

2023 arXiv
[25]

Xiao, Y.; Sun, E.; Luo, D.; and Wang, W. 2024. TradingAgents: Multi-Agents LLM Financial Trading Framework. arXiv preprint arXiv:2412.20138

2024 arXiv
[26]

Xing, F. Z. 2024. HAD: Heterogeneous multi-Agent framework for Financial sentiment analysis. arXiv preprint arXiv:2401.05799

2024 arXiv
[27]

Yang, H.; Zhang, B.; Wang, N.; Guo, C.; Zhang, X.; Lin, L.; Wang, J.; Zhou, T.; Guan, M.; Zhang, R.; and Wang, C. D. 2024. FinRobot: An Open-Source AI Agent Platform for Financial Applications using Large Language Models. arXiv preprint arXiv:2405.14767

2024 arXiv
[28]

Yao, S.; Zhao, J.; Yu, D.; Du, N.; Shafran, I.; Narasimhan, K.; and Cao, Y. 2023. React: Synergizing reasoning and acting in language models. In International Conference on Learning Representations (ICLR)

2023
[29]

W.; Cui, Z.; Liu, R.; Xu, Z.; Zhang, D.; Subbalakshmi, K.; Xiong, G.; He, Y.; Huang, J.; Li, D.; and Xie, Q

Yu, Y.; Yao, Z.; Li, H.; Deng, Z.; Jiang, Y.; Cao, Y.; Chen, Z.; Suchow, J. W.; Cui, Z.; Liu, R.; Xu, Z.; Zhang, D.; Subbalakshmi, K.; Xiong, G.; He, Y.; Huang, J.; Li, D.; and Xie, Q. 2025. FINCON: a synthesized LLM multi-agent system with conceptual verbal reinforcement for ...

2025
[30]

Yuan, H.; Wang, S.; and Guo, J. 2024. Alpha-GPT 2.0: Human-in-the-Loop AI for Quantitative Investment. arXiv preprint arXiv:2402.09746

2024 arXiv
[31]

Zhang, H.; Hua, F.; Xu, C.; Kong, H.; Zuo, R.; and Guo, J. 2024 a . Unveiling the Potential of Sentiment: Can Large Language Models Predict Chinese Stock Price Movements? arXiv preprint arXiv:2306.14222

2024 arXiv
[32]

Zhang, W.; Zhao, L.; Xia, H.; Sun, S.; Sun, J.; Qin, M.; Li, X.; Zhao, Y.; Zhao, Y.; Cai, X.; Zheng, L.; Wang, X.; and An, B. 2024 b . A Multimodal Foundation Agent for Financial Trading: Tool-Augmented, Diversified, and Generalist. In Proceedings of the 30th ACM SIGKDD Confer...

2024
[33]

Zhou, Y.; Liu, H.; Chen, Z.; Tian, Y.; and Chen, B. 2025. GSM-Infinite: How Do your LLMs Behave over Infinitely Increasing Reasoning Complexity and Context Length? In Forty-second International Conference on Machine Learning

2025

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

, " * write output.state after.block = add.period write newline

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

[2] [2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

[3] [3]

Bai, J.; Bai, S.; Chu, Y.; Cui, Z.; Dang, K.; Deng, X.; Fan, Y.; Ge, W.; Han, Y.; Huang, F.; Hui, B.; et al. 2023. QWEN TECHNICAL REPORT. arXiv preprint arXiv:2309.16609

arXiv 2023

[4] [4]

Byrd, D.; Hybinette, M.; and Balch, T. H. 2020. ABIDES: Towards High-Fidelity Multi-Agent Market Simulation. In Proceedings of the 2020 ACM SIGSIM Conference on Principles of Advanced Discrete Simulation, SIGSIM-PADS '20, 11–22. New York, NY, USA: Association for Computing Machinery. ISBN 9781450375924

work page 2020

[5] [5]

Chroma. 2024. Context Rot: How LLMs Degrade with Context. https://research.trychroma.com/context-rot

work page 2024

[6] [6]

DeepSeek-AI; Liu, A.; Feng, B.; Xue, B.; Wang, B.; Wu, B.; Lu, C.; Zhao, C.; Deng, C.; Zhang, C.; Ruan, C.; Dai, D.; Guo, D.; Yang, D.; Chen, D.; Ji, D.; Li, E.; Lin, F.; Dai, F.; Luo, F.; Hao, G.; Chen, G.; Li, G.; Zhang, H.; Bao, H.; Xu, H.; Wang, H.; Zhang, H.; Ding, H.; Xin, H.; Gao, H.; Li, H.; Qu, H.; Cai, J. L.; Liang, J.; Guo, J.; Ni, J.; Li, J.; ...

arXiv 2024

[7] [7]

Ding, H.; Wang, J.; Li, Y.; and Chen, H. 2024. Large Language Model Agent in Financial Trading: A Survey. arXiv preprint arXiv:2408.06361

arXiv 2024

[8] [8]

Engle, R. F. 1982. Autoregressive Conditional Heteroscedasticity with Estimates of the Variance of UK Inflation. Econometrica, 50(4): 987--1008

work page 1982