Pith. sign in

REVIEW 2 major objections 4 minor 31 references

FinDPO: Financial Sentiment Analysis for Algorithmic Trading through Preference Optimization of LLMs

T0 review · 2 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash

Pith's one-line read A preference-tuned financial LLM beats supervised fine-tuning and stays profitable after realistic trading costs.

desk verdict Plausible DPO-for-sentiment idea, but the headline trading returns rely on a same-day backtest that looks look-ahead; the numbers are not credible as written. read the letter →

arxiv 2507.18417 v1 pith:PT4436LQ submitted 2025-07-24 cs.CL cs.LGq-fin.STq-fin.TR

classification cs.CLcs.LGq-fin.STq-fin.TR
keywords financialsentimentanalysisdirectpreferenceoptimizationlargelanguagemodelsalgorithmictradingportfolioconstructionlong-shorttransactioncostslogit-to-scoreconversion
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

This paper argues that supervised fine-tuning, the standard way to adapt large language models to financial sentiment analysis, can be replaced by Direct Preference Optimization (DPO) with better results. The authors introduce FinDPO, a DPO-aligned model, and report that it beats existing fine-tuned models by 11% on average weighted F1 across three financial news benchmarks. They also convert the model's discrete sentiment labels into continuous scores by applying a softmax to the first-token logits and calibrating with temperature scaling, which lets a causal LLM be ranked and traded in a long-short portfolio for the first time. In backtests from 2015 to 2021, FinDPO is the only sentiment method they test that keeps substantial positive performance under 5 basis-point transaction costs: 67% annualized returns and a Sharpe ratio of 2.0. The paper's claim matters because it suggests preference alignment, rather than a larger base model, is what improves generalization in financial language, and that the improvement is large enough to survive real trading frictions.

What carries the argument

The load-bearing mechanism is the pairing of DPO alignment with a 'logit-to-score' converter. DPO replaces the cross-entropy objective of SFT with the preference loss in Equation 1, which increases the probability of the correct sentiment label and decreases the probability of a sampled wrong label relative to the frozen reference model; this explicit penalty is what the paper credits for better generalization to unseen financial text. The logit-to-score converter then takes the logits of the first generated token, applies a softmax over the three sentiment classes, and applies temperature scaling optimized on the training set, because preference-aligned models tend to output degenerate probabilities of 1.0 and 0.0. The resulting continuous, calibrated score is what lets the model rank stocks daily for the long-short portfolio, connecting generative LLM outputs to portfolio weights.

What would settle it

Re-run the long-short backtest using day t sentiment to trade at day t+1's open or close, or restrict the news corpus to articles timestamped before the market's close; if the 5 bps annualized return and Sharpe ratio of 2.0 are no longer achieved, the real-time tradeability claim is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that post-training a finance LLM with DPO rather than SFT produces a sentiment classifier that is both more accurate and more tradeable. On the classification side, FinDPO reports an average weighted F1 of 0.846 over the FPB, TFNS, and NWGI benchmarks, 11% above FinGPT v3.3, and it also outperforms its own instruction-tuned counterpart at 0.771, which the paper takes as evidence that the gain comes from DPO rather than from the base model. On the trading side, the paper claims that by converting the first generated token's logits into a temperature-calibrated probability score, a causal LLM can rank 417 S&P 500 stocks daily, placing the top 35% long and the bottom 35% short. In simulated backtests with daily rebalancing, FinDPO achieves an annualized return of 111.78% with a Sharpe ratio of 3.41 at zero cost, and remains at 66.64% annualized with a Sharpe ratio of 2.03 after 5 bps transaction costs, where all other tested methods turn negative or near-zero. The paper presents this as the first demonstration that a sentiment-based strategy built on a causal LLM can remain profitable under realistic costs.

Load-bearing premise

The backtest assumes that a company's news on day t is available in time to trade that company's day-t return, with no delay between publication, sentiment scoring, and execution.

Editorial extensions

If this is right

  • If FinDPO's classification result holds, DPO becomes a practical substitute for SFT in financial NLP, promising better accuracy and generalization without a larger base model.
  • If the backtest holds up, causal LLMs do not need a classification head to drive portfolios; their generative logits are enough to produce ranking scores.
  • The 5 bps result implies that sentiment signals can be strong enough to pay for realistic trading costs, making LLM-based sentiment a plausible input to systematic strategies.
  • The reported training budget, about 42 million trainable parameters on a single GPU, suggests preference alignment is cheap enough for finance teams to run on their own data.
  • The logit-to-score converter plus temperature scaling gives a reusable recipe for turning any causal LLM's discrete sentiment output into a continuous signal.

Reading between the lines

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

  • The paper leaves the same-day timing assumption untested; a natural next test is to trade day t sentiment at day t+1's open, which would show whether the scores predict future returns or merely track same-day price moves.
  • Because the preference pairs are synthesized from ground-truth labels rather than elicited from human annotators, the recipe could extend to preference data built from model rationales or analyst commentary.
  • The portfolio experiment covers only 2015-2021 S&P 500 constituents; testing the ranking signal on post-2021 data, smaller caps, or international markets would reveal whether the performance is period-specific.
  • Applying the same logit-to-score conversion to other causal LLMs would separate the contribution of DPO alignment from the contribution of score calibration.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. The paper proposes FinDPO, a financial sentiment analysis model based on Direct Preference Optimization (DPO) applied to Llama-3-8B-Instruct, and a 'logit-to-score' converter that maps discrete sentiment labels to continuous scores for long-short portfolio construction. The authors report a weighted F1 of 0.846 on three sentiment benchmarks, claiming an 11% improvement over FinGPT v3.3, and a backtest on S&P 500 news from 2015–2021 that yields an annualized return of 66.64% with a Sharpe ratio of 2.03 at 5 bps transaction costs. The paper also emphasizes parameter-efficient training on a single A100 GPU.

Significance. If the claims hold, FinDPO would be a useful contribution to financial NLP by demonstrating that DPO-based alignment can improve sentiment classification over supervised fine-tuning, and the logit-to-score converter offers a practical way to use causal LLM probabilities in portfolio ranking. The low-resource training setup is also a strength. However, the central trading result rests on a timing assumption that is not stated, and the classification comparison with published baselines is not controlled, so the headline numbers are not currently supported.

major comments (2)
  1. [§4.2, Eqs. (3)–(5)] The backtest aligns day-t news sentiment (Eq. 2) with day-t stock returns (Eqs. 3–5) without specifying any lag between article publication and portfolio execution. If any article used in S_t is published after the market close on day t, the same-day return cannot be realized, and the reported 66.64% annualized return and Sharpe ratio of 2.03 at 5 bps (Table 4) are not achievable. The paper must either provide timestamp statistics demonstrating that all signals are available before the close, or re-run the backtest with a one-day lag and report whether the results survive; without this, the abstract's primary trading claim is unsupported.
  2. [§5.1, Table 2] The comparison with FinGPT v3.3 and other published baselines is not apples-to-apples. FinDPO is trained on 80% of the same three datasets and evaluated on the remaining 20%, while FinGPT v3.3's scores are taken from prior work with unknown training and evaluation splits. Moreover, FinDPO trails FinGPT v3.3 on FPB (0.865 vs 0.879) and TFNS (0.872 vs 0.903), and the claimed 11% average improvement is driven entirely by the NWGI column (0.833 vs 0.643). The 'state-of-the-art by 11%' claim requires a controlled comparison in which all baselines are retrained on the same splits under the same protocol.
minor comments (4)
  1. [Table 1] The header 'No. of articles pre processing' contains a typo; it should read 'pre-processing'.
  2. [§4.2, Portfolio Construction] The text states 'the top 35% of companies in terms of performance were allocated to long positions', but the ranking is based on sentiment scores, not historical performance; this wording should be corrected to avoid confusion.
  3. [§4.2, Eq. (7)] The definition of annualized return R_p as (1/N) Σ r_log(i) · 252 is the annualized log return, not the simple annualized return; the authors should clarify whether the reported values are log or simple returns, as this affects comparability with standard finance metrics.
  4. [§5.3] The paper reports results for a single historical backtest period without statistical significance tests or confidence intervals for the Sharpe ratios; given the autocorrelation of daily returns, a bootstrap or subsample analysis would strengthen the claim of robustness.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found; the central results are empirical backtests and held-out evaluations, with only a non-load-bearing self-citation to FinLlama.

full rationale

FinDPO's claims are empirical rather than derivational. The DPO training uses external ground-truth labels from three public datasets; classification performance is measured on held-out test splits of those datasets, which is standard supervised evaluation, not a circular reduction. The portfolio backtest takes sentiment scores from the model (with softmax and temperature scaling calibrated on the training split) and combines them with market returns; no portfolio return or Sharpe ratio is algebraically forced by the fitted parameter, and the financial article corpus used for portfolio construction is stated to be disjoint from the calibration set. The only self-citation, FinLlama [10], is used as a baseline and for the 35% long/short allocation convention ('Similar to [10]'), but the FinDPO results are independently computed rather than derived from that citation. The paper's main risk is evaluation design: same-day sentiment aligned with same-day returns in Eqs. (2)-(5) without reported timestamps, and the NWGI-inclusive comparison may disadvantage baselines not trained on NWGI. These are correctness and fairness concerns, not circularity; no equation in the paper reduces to its own input.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claims rest on several assumptions that are not independently verified: label-derived preference pairs stand in for human preferences, same-day sentiment is assumed tradeable, benchmark splits are assumed clean, and the softmax score is assumed comparable across companies. These are domain assumptions rather than mathematical axioms, and the paper provides no external validation for them.

free parameters (6)
  • DPO beta = not reported
    Controls the KL penalty in the DPO loss (Eq. 1); no value is given, so the training signal strength is unspecified.
  • Learning rate = not reported
    Described only as 'small'; no numeric value, warmup ratio, or weight decay is provided.
  • LoRA rank = 16
    Chosen by hand; sets the number of trainable parameters to 41.9M (0.52% of the base model).
  • LoRA alpha = 16
    Chosen by hand; scaling factor for LoRA updates.
  • Temperature T = optimized on training set, value not reported
    Fitted by minimizing NLL on the training set to calibrate the logit-to-score probabilities; affects portfolio scores.
  • Long/short allocation = 35%
    Chosen by hand following FinLlama; changes portfolio exposure and turnover.
assumptions (4)
  • domain assumption Preference pairs derived from ground-truth labels and reference-model predictions are a valid proxy for human preferences in DPO.
    Section 4.1.1 constructs pairwise preferences from labels; no human preference annotations are collected, yet the method is described as human preference alignment.
  • domain assumption Sentiment from news on day t is available before the day t return is realized.
    Section 4.2 uses same-day sentiment and returns in Eqs. 3-5; no lag is specified.
  • domain assumption The 80/20 splits of the three training datasets are disjoint and leakage-free.
    Section 4.1.1 states an 80/20 split; the paper does not document the exact split or whether standard benchmark splits were used.
  • domain assumption Softmax over first-token logits yields a meaningful cross-company sentiment strength score.
    Remark 2 defines the conversion, but no calibration or comparability evidence is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FinDPO: Financial Sentiment Analysis for Algorithmic Trading through Preference Optimization of LLMs." pith.science (2026). https://pith.science/paper/PT4436LQ

@misc{pith2026250718417,
  author       = {Pith},
  title        = {Pith review of: FinDPO: Financial Sentiment Analysis for Algorithmic Trading through Preference Optimization of LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PT4436LQ}},
  note         = {Machine review of arXiv:2507.18417}
}
read the original abstract

Opinions expressed in online finance-related textual data are having an increasingly profound impact on trading decisions and market movements. This trend highlights the vital role of sentiment analysis as a tool for quantifying the nature and strength of such opinions. With the rapid development of Generative AI (GenAI), supervised fine-tuned (SFT) large language models (LLMs) have become the de facto standard for financial sentiment analysis. However, the SFT paradigm can lead to memorization of the training data and often fails to generalize to unseen samples. This is a critical limitation in financial domains, where models must adapt to previously unobserved events and the nuanced, domain-specific language of finance. To this end, we introduce FinDPO, the first finance-specific LLM framework based on post-training human preference alignment via Direct Preference Optimization (DPO). The proposed FinDPO achieves state-of-the-art performance on standard sentiment classification benchmarks, outperforming existing supervised fine-tuned models by 11% on the average. Uniquely, the FinDPO framework enables the integration of a fine-tuned causal LLM into realistic portfolio strategies through a novel 'logit-to-score' conversion, which transforms discrete sentiment predictions into continuous, rankable sentiment scores (probabilities). In this way, simulations demonstrate that FinDPO is the first sentiment-based approach to maintain substantial positive returns of 67% annually and strong risk-adjusted performance, as indicated by a Sharpe ratio of 2.0, even under realistic transaction costs of 5 basis points (bps).

Figures

Figures reproduced from arXiv: 2507.18417 by the authors.

Figure 1
Figure 1. FinDPO training principle. Preference pairs are first generated from finance-specific labeled datasets. These preference [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Proposed framework for our sentiment-driven port [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Cumulative returns of sentiment-based long-short portfolios at 0 and 5 bps transaction costs, compared against the [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 21 canonical work pages

  1. [1]

    Dogu Araci. 2019. FinBERT: Financial sentiment analysis with pre-trained language models. arXiv preprint arXiv:1908.10063 (2019)

  2. [2]

    Jacob Boudoukh, Ronen Feldman, Shimon Kogan, and Matthew Richardson. 2013. Which news moves stock prices? A textual analysis . Technical Report. National Bureau of Economic Research

  3. [3]

    Tianzhe Chu, Yuexiang Zhai, Jihan Yang, Shengbang Tong, Saining Xie, Sergey Levine, and Yi Ma. 2025. SFT Memorizes, RL Generalizes: A Comparative Study of Foundation Model Post-training. In The Second Conference on Parsimony and Learning (Recent Spotlight Track)

  4. [4]

    Ronen Feldman, Benjamin Rosenfeld, Roy Bar-Haim, and Moshe Fresko. 2011. The stock sonar—sentiment analysis of stocks based on a hybrid approach. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 25. 1642–1647

  5. [5]

    Sandro Gössi, Ziwei Chen, Wonseong Kim, Bernhard Bermeitinger, and Siegfried Handschuh. 2023. Finbert-fomc: Fine-tuned finbert model with sentiment focus method for enhancing sentiment analysis of fomc minutes. In Proceedings of the Fourth ACM International Conference on AI in Finance . 357–364

  6. [6]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Ab- hishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schel- ten, Alex Vaughan, et al . 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783 (2024)

  7. [7]

    Weinberger

    Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. 2017. On Calibration of Modern Neural Networks. In Proceedings of the 34th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 70) . PMLR, 1321–1330

  8. [8]

    Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-Rank Adaptation of Large Language Models. In The Tenth International Conference on Learning Representa- tions

Show all 31 references
  1. [9]

    Clayton Hutto and Eric Gilbert. 2014. Vader: A parsimonious rule-based model for sentiment analysis of social media text. In Proceedings of the international AAAI conference on web and social media , Vol. 8. 216–225

  2. [10]

    Giorgos Iacovides, Thanos Konstantinidis, Mingxue Xu, and Danilo Mandic. 2024. FinLlama: LLM-Based Financial Sentiment Analysis for Algorithmic Trading. In Proceedings of the 5th ACM International Conference on AI in Finance . 134–141

  3. [11]

    Kelly, and Dacheng Xiu

    Zheng Tracy Ke, Bryan T. Kelly, and Dacheng Xiu. 2019. Predicting Returns With Text Data. NBER Working Papers 26186. National Bureau of Economic Research, Inc. https://EconPapers.repec.org/RePEc:nbr:nberwo:26186

  4. [12]

    Jixuan Leng, Chengsong Huang, Banghua Zhu, and Jiaxin Huang. 2025. Tam- ing Overconfidence in LLMs: Reward Calibration in RLHF. In The Thirteenth International Conference on Learning Representations

  5. [13]

    D. S. Lim. 2021. BERT-base-NER. https://huggingface.co/dslim/bert-base-NER

  6. [14]

    Ilya Loshchilov, Frank Hutter, et al. 2017. Fixing weight decay regularization in adam. arXiv preprint arXiv:1711.05101 5 (2017), 5

  7. [15]

    Tim Loughran and Bill Mcdonald. 2011. When Is a Liability NOT a Liability? Textual Analysis, Dictionaries, and 10-Ks. The Journal of Finance 66 (02 2011), 35 – 65

  8. [16]

    Neural Magic. 2022. Twitter Financial News Sentiment. http://precog.iiitd.edu. in/people/anupama

  9. [17]

    Pekka Malo, Ankur Sinha, Pekka Korhonen, Jyrki Wallenius, and Pyry Takala

  10. [18]

    Long Ouyang et al. 2022. Training language models to follow instructions with human feedback. In Proceedings of the 36th International Conference on Neural Information Processing Systems (New Orleans, LA, USA) (NIPS ’22). Article 2011, 15 pages

  11. [19]

    Yangtuo Peng and Hui Jiang. 2016. Leverage Financial News to Predict Stock Price Movements Using Word Embeddings and Deep Neural Networks. In Pro- ceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Tec...

  12. [20]

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2023. Direct Preference Optimization: Your Language Model is Secretly a Reward Model. In Thirty-seventh Conference on Neural Infor- mation Processing Systems

  13. [21]

    Rajkumar Ramamurthy et al. 2023. Is Reinforcement Learning (Not) for Natural Language Processing: Benchmarks, Baselines, and Building Blocks for Natu- ral Language Policy Optimization. In The Eleventh International Conference on Learning Representations

  14. [22]

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov

  15. [23]

    P. J. Stone, D. C. Dunphy, M. S. Smith, and D. M. Ogilvie. 1966. The General Inquirer: A Computer Approach to Content Analysis . MIT Press

  16. [24]

    Neng Wang, Hongyang Yang, and Christina Dan Wang. 2023. FinGPT: Instruc- tion Tuning Benchmark for Open-Source Large Language Models in Financial Datasets. NeurIPS Workshop on Instruction Tuning and Instruction Following (2023)

  17. [25]

    Oliver Wang. 2023. News with GPT instructions. https://huggingface.co/datasets/ oliverwang15/news_with_gpt_instructions

  18. [26]

    Williams

    Ronald J. Williams. 2004. Simple Statistical Gradient-Following Algorithms for Connectionist Reinforcement Learning. Machine Learning 8 (2004), 229–256

  19. [27]

    Yahoo Finance. 2023. Treasury yield 10 years historical data. https://finance. yahoo.com/quote/%5ETNX/history

  20. [28]

    Boyu Zhang, Hongyang Yang, and Xiao-Yang Liu. 2023. Instruct-FinGPT: Fi- nancial Sentiment Analysis by Instruction Tuning of General-Purpose Large Language Models. ArXiv abs/2306.12659 (2023). https://api.semanticscholar.org/ CorpusID:259224880

  21. [29]

    Wenbin Zhang and Steven Skiena. 2010. Trading Strategies to Exploit Blog and News Sentiment. Proceedings of the International AAAI Conference on Web and Social Media 4, 1 (May 2010), 375–378

  22. [2014]

    Journal of the Association for Information Science and Technology 65, 4 (2014), 782–796

    Good debt or bad debt: Detecting semantic orientations in economic texts. Journal of the Association for Information Science and Technology 65, 4 (2014), 782–796

  23. [2017]

    arXiv preprint arXiv:1707.06347 (2017)

    Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017)

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.