Pith. sign in

REVIEW 4 major objections 4 minor 47 references

This study claims that, in a simulated 12-month backtest, GPT-4 Turbo and FinGPT generate trading signals that outperform a passive S&P 500 benchmark, implying LLMs have measurable predictive content for technical market analysis.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 23:25 UTC pith:CKNAICBM

load-bearing objection The four-task LLM comparison is useful, but the backtest's 'beat the S&P 500' claim collapses once you notice the equal- vs cap-weight mismatch; the paper also reads at points like a commercial plug. the 4 major comments →

arxiv 2607.15414 v1 pith:CKNAICBM submitted 2026-07-16 cs.LG cs.AIq-fin.CP

AI Trading: Evaluating Large Language Models for Technical Market Analysis

Geofrey Ntale This is my paper
classification cs.LG cs.AIq-fin.CP
keywords AI tradinglarge language modelstechnical analysisbacktestingSharpe ratiocandlestick patternsfinancial NLPsignal generation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper evaluates five large language models on four technical-analysis tasks: candlestick pattern recognition, BUY/SELL/HOLD signal generation, backtesting, and financial report comprehension. Its central claim is that the strongest models—GPT-4 Turbo and the finance-tuned FinGPT—produce signals with statistically meaningful predictive content, beating the S&P 500 benchmark in a simulated 2023 backtest (31.2% vs 20.3% annualized return; Sharpe 1.42 vs 0.88). The paper also finds that no model dominates all tasks: FinGPT leads at pattern recognition, GPT-4 Turbo at signals, Claude 3 Opus at low hallucination, Gemini 1.5 Pro at cost and latency. It documents persistent failure modes, especially numerical hallucination and degradation in sideways markets, and concludes that LLMs are useful components in trading pipelines, not autonomous decision-makers.

Core claim

The paper's core discovery is a performance ranking with task-specific heterogeneity. In the simulated backtest, GPT-4 Turbo achieves the highest annualized return (31.2%) and Sharpe ratio (1.42), while FinGPT achieves 28.4% and 1.31, both above the S&P 500 buy-and-hold (20.3%, 0.88). Across all models, signal accuracy on a three-class task ranges from 50.4% to 58.3%, above the 33% random baseline, and the paper claims the differences are statistically significant. The author interprets this as evidence that LLM-generated signals contain real predictive content under controlled conditions, while cautioning that durability across market regimes and after transaction costs remains unproven.

What carries the argument

The central machinery is the four-task evaluation pipeline: structured prompting of LLMs with OHLCV data and technical indicators (RSI, MACD, Bollinger Bands), conversion of model outputs into directional signals, a long-only backtesting engine with fixed transaction costs and equal-weight portfolio construction across 50 S&P 500 constituents, and standard financial metrics (Sharpe, Sortino, drawdown, information coefficient). The load-bearing step is the signal-to-trade mapping: BUY enters at the next open, SELL exits, HOLD maintains position, which isolates signal quality from position sizing.

Load-bearing premise

The backtest compares an equal-weighted portfolio of 50 large-cap stocks against the market-cap-weighted S&P 500 index, with no equal-weight buy-and-hold baseline, so the claimed outperformance may reflect the portfolio's equal-weight tilt rather than the language models' forecasting skill.

What would settle it

Take the same 50 stocks and the same backtest engine, but replace each model's BUY/SELL/HOLD signal with a random or always-HOLD signal matched to the same trade frequency; if the random baseline achieves comparable return and Sharpe over the same 2023 window (or across multiple resampled windows), the models' predictive edge is not real. Alternatively, add an equal-weight buy-and-hold portfolio of the same 50 stocks as a benchmark; if the LLM strategies no longer beat it, the headline outperformance claim fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If these results hold out-of-sample, LLM-based signals could be integrated as one component in modular trading systems, with task routing such as FinGPT for pattern recognition, Claude for verifiable text, and Gemini for cheap screening.
  • Domain-specific fine-tuning of smaller models (FinGPT) can approach frontier-model performance on structured financial tasks at much lower cost, suggesting parameter-efficient adaptation as a practical route.
  • The documented failure modes, numerical hallucination and sideways-market degradation, imply that live deployment should keep a human in the loop or add signal-quality thresholds, as the paper concludes.
  • The paper's claim that an edge above a random three-class baseline is statistically meaningful under the Efficient Market Hypothesis sets up a research agenda: testing durability across regimes, asset classes, and after market impact.
  • A direct corollary is that practitioners should evaluate LLMs on the full workflow, not on classification accuracy alone, because backtested risk-adjusted return is the metric that matters for trading.

Where Pith is reading between the lines

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

  • If an equal-weight buy-and-hold of the same 50 stocks were added as a benchmark, the gap between LLM strategies and the 20.3% S&P 500 return could shrink substantially; the paper does not report this baseline, so the 'beat the market' conclusion is not yet settled.
  • The reported signal accuracy (~50-58%) is not compared against a baseline that matches the trade frequency or class distribution; a strategy that simply predicts HOLD most often might achieve similar accuracy in a sideways year, so the information coefficient would be a more decisive measure.
  • A testable extension would run the same prompts on shuffled or randomly timed OHLCV windows, or compare against a random-signal strategy with the same entry/exit frequency, to isolate whether the LLM's edge comes from price information or from the strategy's long bias in a rising market.
  • The paper's single-year, large-cap, long-only design suggests the natural next experiment is out-of-sample testing in a bear or sideways market, where the reported sideways-regime degradation predicts poor performance.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. This paper compares five LLMs (GPT-4 Turbo, Claude 3 Opus, Gemini 1.5 Pro, Llama 3 70B, FinGPT) on four tasks: candlestick pattern recognition, directional signal generation (BUY/SELL/HOLD), a 12-month simulated backtest (Jan–Dec 2023) on 50 S&P 500 stocks, and financial Q&A benchmarks. The headline findings are that GPT-4 Turbo achieves an annualized return of 31.2% and Sharpe ratio of 1.42, and FinGPT 28.4% and 1.31, both exceeding a passive S&P 500 benchmark (20.3%, 0.88). The paper concludes that LLMs 'can generate signals with statistically meaningful predictive content' and discusses failure modes such as numerical hallucination. The manuscript labels all results as 'simulated/illustrative.'

Significance. The multi-task comparison is potentially useful, and the explicit separation of a 2022 development set for prompt selection from the 2023 evaluation window is a good practice. However, the two central quantitative claims—signal accuracy above chance and backtest outperformance—rest on a mis-specified random baseline and a mismatched benchmark, respectively. As submitted, the evidence does not establish the abstract's conclusion that LLM signals have statistically meaningful predictive content; the main value at this stage is as a framework proposal rather than an empirical demonstration.

major comments (4)
  1. [§3.6 and Table 5] The backtest strategy is a long-only, equal-weighted portfolio of 50 S&P 500 constituents, while the benchmark is the market-cap-weighted S&P 500 index. The paper never reports a buy-and-hold equal-weight baseline for the same 50 stocks. The outperformance of GPT-4 Turbo (31.2% vs 20.3%) and FinGPT (28.4%) could therefore be driven by the equal-weight tilt, the selection of the 50 stocks (whose sampling rule is not described in §3.3), or the timing of entries/exits, rather than by the LLMs' signal content. This is not a minor caveat: without the matched baseline, the central claim that LLMs 'outperform a passive S&P 500 benchmark' is unidentifiable.
  2. [§4.2 and Table 4] The assertion that 'a random baseline would achieve 33% accuracy' assumes a balanced three-class distribution. The ground-truth distribution of BUY/SELL/HOLD under the forward-return rule is not reported and is unlikely to be uniform for daily equity returns. A random classifier that mimics the base rates will achieve roughly the largest class frequency, so accuracy values of 50–58% are not 'substantially above chance' unless compared with the correct baseline. Please report class frequencies, per-class precision/recall, and a baseline such as always-majority-class or a random classifier with the empirical class proportions.
  3. [§4.3 and Table 5] The 12-month backtest is a single historical path with no measure of uncertainty. No confidence intervals, bootstrap, or sub-period analyses are provided, and the five-model comparison is subject to multiple-comparison issues. The claim that GPT-4 Turbo outperforms the S&P 500 (Sharpe 1.42 vs 0.88) could be within sampling variation. Provide at least block-bootstrap standard errors or a paired test across the 50 equities, and report the test statistics for the McNemar test mentioned in §4.2.
  4. [Reproducibility (throughout)] Section 3.2 states that prompt templates are 'documented in the supplementary materials,' and §3.6 describes the pipeline, but no code, data, stock list, prompt templates, or evaluation logs are provided. The 50-stock selection rule is also absent. Without these artifacts, the numerical results cannot be verified and the reported significance claims cannot be audited. Please release the evaluation artifacts (including prompts, model outputs, and the universe construction script) under a suitable license.
minor comments (4)
  1. [Table captions and §4] The repeated label 'Simulated / Illustrative Data' creates ambiguity about whether the numbers are actual experimental measurements or hypothetical examples. If the OHLCV data are historical and the results are real experimental outputs, say so explicitly; if they are illustrative, this must be clearly stated in the abstract and conclusion.
  2. [Table 5 footnote] The risk-free rate is stated as 5.25% (annualized 3-month T-bill yield, Jan 2023). Historical data for January 2023 show the 3-month T-bill at approximately 4.6–4.7%. Please correct this value or provide a source; it affects the reported Sharpe and Sortino ratios.
  3. [§4.4 and Table 6] For FinQA and TAT-QA, a stratified sample of 500 questions is used, but no sampling seed, stratum sizes, or confidence intervals are given. Exact-match accuracy on a 500-question sample is not directly comparable to published full-set scores without a margin of error.
  4. [General presentation] There are several formatting issues: inconsistent capitalization of 'ai trading' vs 'AI trading'; garbled equation renderings in §3.5 (e.g., the Accuracy formula appears as 'Accuracy= TP+TN+FP+FN TP+TN'); and the reference to Fama [30] in §2.4/§5.1 appears misplaced. Also, the repeated promotion of TradieCapital (§3.6 and §5.3) reads as commercial content and should be removed or moved to a footnote.

Circularity Check

0 steps flagged

No significant circularity; the evaluation is empirical, forward-looking, and not derived from its own inputs.

full rationale

The paper does not contain a derivation chain in which an output is defined as its own input. The central backtest result (Table 5) is produced by a pipeline where each LLM receives a rolling 30-day OHLCV window plus pre-computed indicators available at market close, generates a BUY/SELL/HOLD signal, and the signal is then executed at the next day's open and compared against realized forward returns over Jan-Dec 2023. This is an out-of-sample evaluation relative to the signals, not a fitted prediction. No parameter is calibrated to the reported returns, no uniqueness theorem is imported from the authors' prior work, and no self-citation is load-bearing (the reference list contains no works by the present author). Task D is benchmarked against external, publicly released test sets (FinQA, TAT-QA, FinanceBench), which provides independent grounding. The main validity concern -- comparing an equal-weighted 50-stock long-only strategy to a cap-weighted S&P 500 index -- is a real benchmark-matching flaw, but it is a test-design issue, not circularity: even a flawed comparison can be non-circular. The paper also repeatedly labels its results as 'simulated / illustrative' and lists limitations (single market regime, large-cap universe, idealized execution), further indicating that the claims are empirical estimates rather than tautological derivations. Therefore no circular step can be exhibited, and the appropriate score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central claim depends on several hand-picked thresholds and domain assumptions. No free parameters were fit to data; the paper is an empirical evaluation, so its 'free parameters' are methodological choices. The most consequential assumptions are the benchmark comparability and the 33% random baseline.

free parameters (4)
  • Transaction cost = 0.05% per trade (Task B), 0.10% round-trip (Task C)
    Hand-chosen; affects whether signals count as correct and the magnitude of backtest returns.
  • Trending/sideways threshold = 1% per week
    Arbitrary cutoff used to compute regime-dependent signal accuracy in Section 4.5; no sensitivity analysis.
  • HOLD scoring rule = Sign of forward return
    The paper defines a correct signal as matching the sign of next-day return, but does not specify how HOLD is scored; this ambiguity affects Task B accuracy.
  • Prompt template = Not disclosed
    System prompts vary only by model, but templates are only said to be in supplementary materials; prompts affect outputs.
axioms (4)
  • domain assumption Technical patterns (candlesticks, RSI, MACD, Bollinger Bands) carry predictive information for next-day returns
    The evaluation framework assumes technical signals have predictive content, an assumption contested in academic literature; the paper does not test this prior.
  • domain assumption The equal-weighted 50-stock portfolio can be benchmarked against the cap-weighted S&P 500 index
    Used in Table 5 to conclude LLM strategies 'outperform' the market; this is a load-bearing comparison assumption.
  • domain assumption A uniform random baseline of 33% is the appropriate chance level for a three-class BUY/SELL/HOLD task
    Assumed in Section 4.2 when claiming models perform 'substantially above chance'; ignores class imbalance and base rates.
  • ad hoc to paper The provided OHLCV data and prompts contain no look-ahead bias
    Stated in Section 3.6; depends on the implementation being correct, but no code is provided to verify.

pith-pipeline@v1.3.0-alltime-deepseek · 17819 in / 14350 out tokens · 135698 ms · 2026-08-01T23:25:55.235119+00:00 · methodology

0 comments
read the original abstract

Large Language Models (LLMs) have emerged as powerful tools for processing the heterogeneous information environments of modern financial markets. This paper presents a systematic, comparative evaluation of five prominent LLMs: GPT-4 Turbo, Claude 3 Opus, Gemini 1.5 Pro, Llama 3 70B, and the domain-specialized FinGPT, with respect to their capacity for technical market analysis. The evaluation spans four structured tasks: candlestick pattern recognition from OHLCV data, directional signal generation (BUY/SELL/HOLD), backtesting of signal quality through a simulated execution pipeline, and financial report comprehension. Our experimental framework employs rigorous quantitative metrics, including Sharpe ratio, maximum drawdown, Sortino ratio, information coefficient, F1-score, and BLEU score. Findings from simulated backtesting indicate that GPT-4 Turbo achieves the highest annualized return and Sharpe ratio among general-purpose models, while FinGPT demonstrates competitive risk-adjusted performance due to domain-specific fine-tuning. Both models outperform a passive S&P 500 benchmark under the tested conditions. The study identifies persistent failure modes across all evaluated models, including numerical hallucination, context-window limitations, and inconsistent performance in sideways market regimes. We conclude that while LLMs hold genuine promise within AI trading systems, robust deployment requires careful task decomposition, rigorous backtesting protocols, and domain-aware fine-tuning strategies.

discussion (0)

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

Reference graph

Works this paper leans on

47 extracted references · 13 linked inside Pith

  1. [1]

    Analyzing and modelling 1+1d markets,

    D. Challet and R. Stinchcombe, "Analyzing and modelling 1+1d markets," Physica A: Statistical Mechanics and its Applications, vol. 300, no. 1-2, pp. 285-299, 2001

    2001

  2. [2]

    Price formation and optimal trading in intraday electricity markets,

    A. Karpf, A. Mandel, and S. Battiston, "Price formation and optimal trading in intraday electricity markets," Journal of Economic Dynamics and Control, vol. 83, pp. 147-172, 2017

    2017

  3. [3]

    Deep neural networks, gradient-boosted trees, random forests: Statistical arbitrage on the S&P 500,

    C. Krauss, X. A. Do, and N. Huck, "Deep neural networks, gradient-boosted trees, random forests: Statistical arbitrage on the S&P 500," European Journal of Operational Research, vol. 259, no. 2, pp. 689-702, 2017

    2017

  4. [4]

    Text-based sentiment analysis in finance: Synthesising the existing literature and exploring future directions,

    P. Todd, C. Bowden, and Y. Moshfeghi, "Text-based sentiment analysis in finance: Synthesising the existing literature and exploring future directions," Intelligent Systems in Accounting, Finance and Management, vol. 31, no. 1, e1549, 2024

    2024

  5. [5]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, "Attention is all you need," in Advances in Neural Information Processing Systems (NeurIPS), vol. 30, pp. 5998-6008, 2017

    2017

  6. [6]

    Bulkowski, Encyclopedia of Candlestick Charts

    T. Bulkowski, Encyclopedia of Candlestick Charts. Hoboken, NJ: John Wiley & Sons, 2008

    2008

  7. [7]

    Deep learning techniques in algorithmic trading: A comprehensive review,

    Y. Chen, Y. Lin, and X. Chen, "Deep learning techniques in algorithmic trading: A comprehensive review," arXiv preprint, arXiv:2502.15853, 2025

    Pith/arXiv arXiv 2025

  8. [8]

    Long short-term memory,

    S. Hochreiter and J. Schmidhuber, "Long short-term memory," Neural Computation, vol. 9, no. 8, pp. 1735- 1780, 1997

    1997

  9. [9]

    A survey of large language models in finance: From foundations to frontiers,

    H. Yang, X. Liu, C. Wang, and D. Wang, "A survey of large language models in finance: From foundations to frontiers," arXiv preprint, arXiv:2507.01990, 2025

    Pith/arXiv arXiv 2025

  10. [10]

    Giving content to investor sentiment: The role of media in the stock market,

    P. C. Tetlock, "Giving content to investor sentiment: The role of media in the stock market," The Journal of Finance, vol. 62, no. 3, pp. 1139-1168, 2007

    2007

  11. [11]

    When is a liability not a liability? Textual analysis, dictionaries, and 10-Ks,

    T. Loughran and B. McDonald, "When is a liability not a liability? Textual analysis, dictionaries, and 10-Ks," The Journal of Finance, vol. 66, no. 1, pp. 35-65, 2011

    2011

  12. [12]

    FinBERT: Financial sentiment analysis with pre-trained language models,

    D. Araci, "FinBERT: Financial sentiment analysis with pre-trained language models," arXiv preprint, arXiv:1908.10063, 2019

    Pith/arXiv arXiv 1908

  13. [13]

    GPT-4 Technical Report,

    OpenAI, "GPT-4 Technical Report," arXiv preprint, arXiv:2303.08774, 2023

    Pith/arXiv arXiv 2023

  14. [14]

    Claude 3 Model Card,

    Anthropic, "Claude 3 Model Card," Anthropic Technical Report, 2024. [Online]. Available: https://www.anthropic.com/

    2024

  15. [15]

    Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context,

    Google DeepMind, "Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context," arXiv preprint, arXiv:2403.05530, 2024

    Pith/arXiv arXiv 2024

  16. [16]

    Llama 3: Meta's most advanced publicly available frontier language model,

    Meta AI, "Llama 3: Meta's most advanced publicly available frontier language model," Meta AI Blog, Apr

  17. [17]

    FinGPT: Open-source financial large language models,

    H. Yang, X. Liu, and C. D. Wang, "FinGPT: Open-source financial large language models," arXiv preprint, arXiv:2306.06031, 2023. 30

    arXiv 2023

  18. [18]

    J. J. Murphy, Technical Analysis of the Financial Markets: A Comprehensive Guide to Trading Methods and Applications. New York: New York Institute of Finance, 1999

    1999

  19. [19]

    Nison, Japanese Candlestick Charting Techniques: A Contemporary Guide to the Ancient Investment Techniques of the Far East, 2nd ed

    S. Nison, Japanese Candlestick Charting Techniques: A Contemporary Guide to the Ancient Investment Techniques of the Far East, 2nd ed. New York: Prentice Hall Press, 2001

    2001

  20. [20]

    BloombergGPT: A large language model for finance,

    S. Wu, O. Irsoy, S. Lu, V. Dabravolski, M. Dredze, S. Gehrmann, P. Kambadur, D. Rosenberg, and G. Mann, "BloombergGPT: A large language model for finance," arXiv preprint, arXiv:2303.17564, 2023

    Pith/arXiv arXiv 2023

  21. [21]

    FinQA: A dataset of numerical reasoning over financial data,

    Z. Chen, W. Chen, C. Smiley, S. Shah, I. Borova, D. Langdon, R. Moussa, M. Beane, T.-H. Huang, B. Routledge, and W. Wang, "FinQA: A dataset of numerical reasoning over financial data," in Proc. EMNLP, pp. 3670-3685, 2021

    2021

  22. [22]

    TAT-QA: A question answering benchmark on a hybrid of tabular and textual content in finance,

    F. Zhu, W. Lei, Y. Huang, C. Wang, S. Zhang, J. Lv, F. Feng, and T.-S. Chua, "TAT-QA: A question answering benchmark on a hybrid of tabular and textual content in finance," in Proc. ACL, pp. 3277-3287, 2021

    2021

  23. [23]

    FinanceBench: A new benchmark for financial question answering,

    P. Islam, M. H. Kannappan, D. Kiela, R. Long, N. Patel, and W. Ternes, "FinanceBench: A new benchmark for financial question answering," arXiv preprint, arXiv:2311.11944, 2023

    Pith/arXiv arXiv 2023

  24. [24]

    FinBen: A holistic financial bench‐ mark for large language models,

    Q. Xie, W. Han, X. Zhang, Y. Lai, M. Peng, A. Lopez-Lira, and J. Huang, "FinBen: A holistic financial bench‐ mark for large language models," in Advances in Neural Information Processing Systems (NeurIPS), 2024

    2024

  25. [25]

    Can ChatGPT forecast stock price movements? Return predictability and large language models,

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

    arXiv 2023

  26. [26]

    Learning to generate explainable stock predictions using self-reflec‐ tive large language models,

    K. Koa, Y. Ma, R. Bing, and T.-S. Chua, "Learning to generate explainable stock predictions using self-reflec‐ tive large language models," in Proc. WWW, pp. 4304-4315, 2024

    2024

  27. [27]

    Finance agent benchmark: A human-in-the-loop evaluation harness for LLM agents in the finance domain,

    J. Schmitt, P. Wollstadt, and T. Döring, "Finance agent benchmark: A human-in-the-loop evaluation harness for LLM agents in the finance domain," arXiv preprint, arXiv:2508.00828, 2025

    Pith/arXiv arXiv 2025

  28. [28]

    Temporal data meets LLM: Explainable financial time series forecasting,

    Y. Yu, F. Li, and Y. Chen, "Temporal data meets LLM: Explainable financial time series forecasting," arXiv preprint, arXiv:2306.11025, 2023

    Pith/arXiv arXiv 2023

  29. [29]

    FinMem: A performance-enhanced LLM trading agent with layered memory and character design,

    W. Jin, H. Cao, A. Yang, J. Zhang, and M. McKeown, "FinMem: A performance-enhanced LLM trading agent with layered memory and character design," arXiv preprint, arXiv:2311.13743, 2023

    Pith/arXiv arXiv 2023

  30. [30]

    Efficient capital markets: A review of theory and empirical work,

    R. Fama, "Efficient capital markets: A review of theory and empirical work," Journal of Finance, vol. 25, no. 2, pp. 383-417, 1970

    1970

  31. [31]

    Active retrieval augmented generation,

    Z. Jiang, F. Xu, L. Gao, Z. Sun, Q. Liu, J. Dwivedi-Yu, Y. Yang, J. Callan, and G. Neubig, "Active retrieval augmented generation," in Proc. EMNLP, 2023

    2023

  32. [32]

    A backtesting protocol in the era of machine learning,

    R. D. Arnott, C. R. Harvey, and H. Markowitz, "A backtesting protocol in the era of machine learning," The Journal of Financial Data Science, vol. 1, no. 1, pp. 64-74, 2019

    2019

  33. [33]

    and the cross-section of expected returns,

    C. R. Harvey, Y. Liu, and H. Zhu, "...and the cross-section of expected returns," Review of Financial Studies, vol. 29, no. 1, pp. 5-68, 2016

    2016

  34. [34]

    Survey of hallucina‐ tion in natural language generation,

    Z. Ji, N. Lee, R. Frieske, T. Yu, D. Su, Y. Xu, E. Ishii, Y. Bang, A. Madotto, and P. Fung, "Survey of hallucina‐ tion in natural language generation," ACM Computing Surveys, vol. 55, no. 12, pp. 1-38, 2023. 31

    2023

  35. [35]

    Lost in the middle: How language models use long contexts,

    N. F. Liu, K. Lin, J. Hewitt, A. Paranjape, M. Bevilacqua, F. Petroni, and P. Liang, "Lost in the middle: How language models use long contexts," Transactions of the Association for Computational Linguistics, vol. 12, pp. 157-173, 2024

    2024

  36. [36]

    Self-refine: Iterative refinement with self-feedback,

    A. Madaan, N. Tandon, P. Gupta, S. Hallinan, L. Gao, S. Wiegreffe, U. Alon, N. Dziri, S. Prabhumoye, Y. Yang, S. Welleck, and H. Hajishirzi, "Self-refine: Iterative refinement with self-feedback," in Advances in Neural Information Processing Systems (NeurIPS), 2023

    2023

  37. [37]

    Market efficiency, long-term returns, and behavioral finance,

    E. F. Fama, "Market efficiency, long-term returns, and behavioral finance," Journal of Financial Economics, vol. 49, no. 3, pp. 283-306, 1998

    1998

  38. [38]

    LLM-based agents for trading: A comparative study,

    D. Kim, M. Kim, T. Kim, and K. Park, "LLM-based agents for trading: A comparative study," in Proc. AAAI Workshop on AI in Finance, 2024

    2024

  39. [39]

    AI-Powered Trading and Market Intelligence Platform,

    TradieCapital, "AI-Powered Trading and Market Intelligence Platform," tradiecapital.com, 2024. [Online]. Available: https://tradiecapital.com

    2024

  40. [40]

    Risk Alert: Observations from Examinations of Investment Advisers Using Automated Investment Tools,

    U.S. Securities and Exchange Commission, "Risk Alert: Observations from Examinations of Investment Advisers Using Automated Investment Tools," SEC, 2021

    2021

  41. [41]

    Regulation (EU) 2016/679 (MiFID II),

    European Parliament and Council, "Regulation (EU) 2016/679 (MiFID II)," Official Journal of the European Union, 2018

    2016

  42. [42]

    The limits of arbitrage,

    A. Shleifer and R. W. Vishny, "The limits of arbitrage," The Journal of Finance, vol. 52, no. 1, pp. 35-55, 1997

    1997

  43. [43]

    What happened to the quants in August 2007? Evidence from factors and transac‐ tions data,

    M. Khandani and A. W. Lo, "What happened to the quants in August 2007? Evidence from factors and transac‐ tions data," Journal of Financial Markets, vol. 14, no. 1, pp. 1-46, 2011

    2007

  44. [44]

    MME-Finance: A multimodal finance benchmark for expert-level understanding and reasoning,

    Z. Duan, H. Hua, Z. Dong, J. Liu, J. Xu, X. Liao, Z. Peng, W. Hao, Z. Guo, and X. Kong, "MME-Finance: A multimodal finance benchmark for expert-level understanding and reasoning," arXiv preprint, arXiv:2411.03314, 2024

    Pith/arXiv arXiv 2024

  45. [45]

    FinChart-Bench: A benchmark for financial chart-based comprehension and analysis,

    J. Guo et al., "FinChart-Bench: A benchmark for financial chart-based comprehension and analysis," arXiv preprint, arXiv:2507.14823, 2025

    Pith/arXiv arXiv 2025

  46. [46]

    FinRL: A deep reinforcement learning library for automated stock trading in quantitative finance,

    X. Liu, H. Yang, J. Gao, and C. D. Wang, "FinRL: A deep reinforcement learning library for automated stock trading in quantitative finance," arXiv preprint, arXiv:2011.09607, 2021

    Pith/arXiv arXiv 2011

  47. [2024]

    Available: https://ai.meta.com/

    [Online]. Available: https://ai.meta.com/