FinSMART: Financial Sentiment Analysis for Algorithmic Trading through Market-Aligned Reinforcement Learning

Danilo Mandic; Giorgos Iacovides; Wuyang Zhou

REVIEW 4 major objections 5 minor 20 references

Financial sentiment models can be trained directly on market returns instead of human labels, and the resulting signals beat prior methods on live-style trading tests.

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 · grok-4.5

2026-07-31 17:00 UTC pith:5K5JI6XZ

load-bearing objection Solid applied step—GRPO on market outcomes beats static-label finance LLMs in one backtest—but same-day rewards vs next-day trading and an abstract/body return mismatch keep the edge provisional. the 4 major comments →

arxiv 2607.28127 v1 pith:5K5JI6XZ submitted 2026-07-30 cs.CL cs.LGq-fin.STq-fin.TR

FinSMART: Financial Sentiment Analysis for Algorithmic Trading through Market-Aligned Reinforcement Learning

Giorgos Iacovides , Wuyang Zhou , Danilo Mandic This is my paper

classification cs.CL cs.LGq-fin.STq-fin.TR

keywords financial sentiment analysisreinforcement learningGRPOalgorithmic tradinglarge language modelsmarket feedbacklong-short portfolios

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.

This paper argues that financial sentiment analysis has been stuck in a dead end: models are fine-tuned on small, static, human-labelled datasets that never see what the market actually did afterward. FinSMART closes that loop by treating realized stock returns as the reward and training a language model with reinforcement learning so its positive, negative, or neutral calls are scored by whether they would have made money. A filtering pipeline and an asymmetric trading reward keep the noisy market signal usable. On out-of-sample long–short portfolios the method substantially lifts cumulative returns, risk-adjusted ratios, and cross-sectional predictive power over the strongest prior systems, and periodic retraining on fresh articles plus their outcomes improves results further. A sympathetic reader cares because this replaces expensive annotation with automatic market feedback and offers a path to sentiment models that keep adapting as regimes change.

Core claim

FinSMART shows that a pre-trained language model can be aligned to financial sentiment by optimizing a group-relative policy objective against a dual-filter asymmetric trading reward built from publication-day raw and alpha returns, and that the resulting continuous sentiment scores produce long–short portfolios with markedly higher cumulative return, Sharpe, Sortino, Calmar, and RankIC than lexicon, supervised, and preference-optimized baselines, with further gains from six-month market-aligned retraining.

What carries the argument

The dual-filter asymmetric trading reward: a prediction earns a large positive score only when its direction matches a same-day move that is both profitable and above a minimum idiosyncratic-alpha threshold; wrong direction, false long/short, and neutral cases receive smaller or negative scores. Group Relative Policy Optimization then ranks multiple sampled labels per article and updates the policy under a KL constraint, so learning rides relative economic ranking rather than absolute noisy returns.

Load-bearing premise

The load-bearing premise is that training on same-day market returns teaches a sentiment signal that still predicts next-day open-to-open trading profits rather than merely echoing contemporaneous price co-movement.

What would settle it

Retrain with the identical pipeline but replace publication-day rewards with next-day rewards only; if the out-of-sample long–short edge and RankIC collapse to baseline levels, the central claim that market-aligned RL yields transferable predictive sentiment is falsified.

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

If this is right

Sentiment LLMs can be updated whenever new articles and returns arrive, without fresh human annotation.
Portfolio construction can draw continuous scores from a causal LLM whose discrete labels were optimized for economic payoff, not label accuracy alone.
Market-aware filtering plus asymmetric rewards become a reusable recipe for stable RL on other noisy financial targets.
Static labelled finance datasets become optional rather than mandatory for post-training sentiment models.

Where Pith is reading between the lines

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

The same reward-and-filter pattern could be tried on earnings transcripts or filings where the economic outcome is a multi-day drift rather than a single open-to-open move.
If same-day training mainly captures liquidity or attention effects, shorter-horizon microstructure features might explain part of the reported edge and should be ablated in follow-ups.
Expanding the investable universe beyond S&P 500 names would test whether the learned alignment generalizes to less-covered equities where news is scarcer and noisier.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

Solid applied step—GRPO on market outcomes beats static-label finance LLMs in one backtest—but same-day rewards vs next-day trading and an abstract/body return mismatch keep the edge provisional.

read the letter

Punchline: this is a real methodological move in finance NLP—train a causal LLM with GRPO on filtered news and a dual-filter asymmetric market reward instead of human sentiment labels—and the 2019–2021 long–short numbers beat FinDPO and the usual baselines. The expanding-window retrain is the part that actually feels useful in practice.

What is new is not GRPO itself, but closing the loop: NER + sentiment gate, y from publication-day alpha/raw with τ=0.5%, asymmetric R_trade, LoRA GRPO on Llama-3-8B, then logit-to-score portfolios. Chronological split, next-day open-to-open evaluation, RankIC, and static vs six-month retrain (265% → 406% cum) are done cleanly. Citations track the FinBERT/FinGPT/FinLlama/FinDPO line honestly. Math is standard GRPO plus a discrete reward; nothing exotic, nothing broken.

Soft spots, in proportion. The load-bearing one is Remark 2 / §4.2: they train on same-day returns because same-day text–return separation is huge (~5%/2.3% alpha, r≈0.4) and collapses one day later (~0.3%, r≈0.03). Trading metrics are next-day. That is intentional SNR engineering, not look-ahead in the backtest, but it leaves open whether the policy learned transferable predictive content or contemporaneous co-movement. Until someone shows reward-horizon ablations or multi-period costed tests, treat the economic edge as conditional. Second: abstract claims +220% over strongest baseline; body is 264.9 vs 109.8 ≈ +141%. That is sloppy overclaim on the headline number. Minor: no costs, single window, free knobs (τ, reward magnitudes, G, β, 35% L/S), no public code/data.

Who it is for: people building event-driven or sentiment portfolios and anyone post-training finance LLMs without labels. Not a theory paper. I would bring it to reading group as a methods + backtest discussion, not as settled alpha. It deserves a serious referee—important enough and formally clear enough—even if revision should force the same-day/next-day issue and fix the abstract. Engage with the work; do not treat the P&L as gospel until replication.

Referee Report

4 major / 5 minor

Summary. The paper proposes FinSMART, a market-aligned RL framework that post-trains a causal LLM (Llama-3-8B-Instruct via LoRA+GRPO) for financial sentiment analysis by optimizing a dual-filter asymmetric trading reward derived from publication-day raw and idiosyncratic returns, rather than static human labels. A three-stage pipeline (NER entity linking, reference-model sentiment gating, and GRPO with group-relative advantages) is used to extract a usable learning signal from noisy markets. Out-of-sample long–short (35%) portfolios on 2019–mid-2021 news show large gains over lexicon, FinBERT, FinLlama, and FinDPO baselines (static FinSMART: 264.9% cumulative return, Sharpe 1.97, RankIC 0.061 vs FinDPO 109.8%, 1.12, 0.053), and six-month expanding-window retraining further lifts cumulative return to 406.2%. The central claim is that market-outcome RL yields more economically informative sentiment and enables continual adaptation without manual annotation.

Significance. If the results hold under stricter predictive controls, this is a meaningful paradigm shift for financial NLP: replacing scarce static labels with realized market feedback, and demonstrating practical continual retraining. Strengths include a clean chronological train/test split, next-day open-to-open backtests that avoid look-ahead in evaluation, NER and gating for reward hygiene, full risk-adjusted metrics plus RankIC, and an explicit static-vs-retrained comparison. Parameter-efficient GRPO on a single A6000 is also a practical contribution. The work is significant for algorithmic trading and adaptive financial LLMs, provided the same-day reward → next-day edge transfer is substantiated.

major comments (4)

[§4.2 Dual-Filter Trading Reward; Remark 2] §4.2 and Remark 2: Training rewards use publication-day raw and alpha returns (τ=0.5%), while all portfolio metrics (§5, Eqs. 5–7, Table 1, Fig. 4) use next-day open-to-open returns. The paper’s own diagnostics show same-day sentiment–alpha separation (~5.0%/2.3%, Pearson 0.41/0.37) collapses under a one-day shift (~0.3%, 0.03). This is load-bearing: if GRPO mainly imprints contemporaneous co-movement or already-priced news rather than transferable predictive content, the claimed next-day trading edge is overstated. Please add (i) an ablation that trains with next-day (or open-to-open) rewards under the same pipeline, (ii) a same-day vs next-day RankIC/return comparison for FinSMART itself, and (iii) discussion of microstructure/timing (e.g., whether articles are available before the close used in same-day r).
[Abstract; Table 1] Abstract vs Table 1: the abstract states a 220% improvement in cumulative trading returns over the strongest baseline, but Table 1 gives FinSMART 264.9% vs FinDPO 109.8%, i.e. a ~141% relative improvement ((264.9−109.8)/109.8). This numerical mismatch must be corrected and the abstract claim aligned with the body; overstated headline numbers undermine confidence in the central empirical claim.
[§4.2; §5 Portfolio Construction; Table 1] §4.2 reward design and free parameters: R_trade assigns +2.0 / +0.1 / −1.5 / −1.0 with τ=0.5%, and GRPO uses G=8, β=0.1. Portfolio construction fixes a 35% long–short cut. There are no ablations on τ, reward asymmetry, neutral reward, or long–short fraction, yet these choices directly define the optimization target and the reported edge. At minimum, sensitivity of Table 1 metrics (and RankIC) to τ and to a symmetric classification-style reward should be reported so the gains are not tied to an untested reward schedule.
[Table 1; §6.1] Table 1: RankIC rises only modestly (0.053 → 0.061, ~15%) while cumulative return more than doubles and Sharpe nearly doubles. That gap suggests concentration, turnover, or tail positioning rather than uniformly better cross-sectional ranking. Please report turnover, average |position| concentration, long/short leg returns separately, and preferably bootstrap or Newey–West significance on Sharpe/RankIC differences versus FinDPO so the economic claim is tied to a clearly stronger predictive signal.

minor comments (5)

[Abstract] Typo in abstract/intro: “mutlifactorial” → “multifactorial”.
[Figure 2; §4.1] Figure 2 caption and §4.1: clarify whether the behavior policy for sampling is π_θold or the frozen reference, and whether sentiment gating is applied only at data construction or also at inference.
[§5 Sentiment Analysis] §5: state explicitly how continuous scores are obtained for FinSMART (logit-to-score from FinDPO) and whether temperature/decoding differs from FinDPO at inference, for reproducibility.
[§6.2; Table 2] Table 2 and right panel of Figure 4: report the exact retraining dates and number of new articles per window; the r=0.72 correlation with article count is interesting but needs sample size and a brief caveat (n≈4 windows).
[§2 Related Work] Related work: briefly situate against other market-feedback or RL-for-trading LLM lines (if any post-FinDPO) so the “first” claim is scoped precisely to sentiment-label-free market-aligned GRPO for FSA.

Circularity Check

0 steps flagged

No significant circularity: market-outcome RL with temporal split and next-day evaluation is intentional alignment, not a by-construction tautology.

full rationale

FinSMART is an empirical RL systems paper, not a first-principles derivation. The load-bearing chain is: (i) filter articles and define a discrete dual-filter reward R_trade(d,y) from publication-day raw/alpha returns (§4.1–4.2); (ii) optimize a LoRA policy with GRPO on the 2015–2018 train split (§4.3); (iii) convert logits to scores and build next-day open-to-open 35% long–short portfolios on 2019–2021 OOS data (§5, Eqs. 5–7), reporting cumulative/annualized return, Sharpe/Sortino/Calmar, and RankIC against external baselines (Table 1, Fig. 4). Training reward and evaluation metric are deliberately related (both economic) but are not the same quantity on the same sample: rewards use same-day α/r with τ=0.5%, while all reported trading metrics use next-day returns and a chronological holdout. That design choice raises a validity/leakage question (same-day SNR vs next-day transfer), not definitional circularity. Reward weights and τ are author-chosen hyperparameters, not parameters fitted to the OOS metrics and then relabeled as predictions. Self-citations to FinLlama/FinDPO supply baselines and the logit-to-score converter; they do not underwrite a uniqueness theorem or force the reported edge. No self-definitional identity, fitted-input-as-prediction, or renamed known law appears in the equations. Abstract vs body return-improvement arithmetic is an overclaim inconsistency, not circularity. Score 0; steps empty.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 2 invented entities

The central performance claim rests on a stack of domain modeling choices (what counts as a tradable sentiment event, how market feedback is turned into a scalar reward, how portfolios are formed) plus standard RL/LLM tooling. Few of these are derived; most are stipulated. Free parameters in the reward and filters directly shape which articles and which market moves teach the policy, so the ledger is dominated by hand-set thresholds rather than invented physical entities.

free parameters (7)

alpha threshold τ = 0.5%
Minimum idiosyncratic move required to label a non-neutral market state y; set to 0.5% throughout without reported sweep tied to the final claim.
asymmetric trading reward magnitudes = +2.0, +0.1, -1.5, -1.0
Discrete rewards +2.0 / +0.1 / −1.5 / −1.0 for correct direction, correct neutral, opposite, and other; chosen to encourage exploration and avoid neutral collapse.
GRPO group size G and KL weight β = G=8, β=0.1
Control exploration and deviation from the reference policy during RL; fixed at G=8, β=0.1.
NER confidence cutoff = 98%
Articles kept only if target-company NER confidence exceeds 98%; directly determines training pairs.
long–short portfolio fraction = 35%
Top/bottom 35% equal-weight book used for all method comparisons, following prior FinLlama/FinDPO setup; headline returns depend on this rule.
LoRA rank and scaling = r=16, α=32, dropout=0.05
r=16, α=32, dropout 0.05 fix the adaptable subspace and effective step size of post-training.
retraining interval = 6 months
Six-month expanding-window updates define the adaptive-FinSMART result (406% cumulative).

axioms (6)

ad hoc to paper Publication-day stock and alpha returns are a higher-SNR proxy of article-implied sentiment than next-day returns and are therefore legitimate RL targets even when evaluation trades next-day.
Stated in §4.2 and Remark 2 with correlation/spread comparisons; load-bearing for the training design.
domain assumption Idiosyncratic alpha α = r_stock − r_SPY plus a sign constraint on raw return isolates stock-specific information attributable to the article rather than market beta.
Used to define ground-truth direction y in the dual-filter reward (§4.2).
ad hoc to paper Reference-model sentiment gating (output restricted to {Positive, Neutral, Negative}) removes undecodable text without discarding economically useful signal.
§4.1 Sentiment Gating; shapes the training distribution.
domain assumption GRPO group-relative advantages plus KL to a frozen reference yield stable policy updates under noisy market rewards without a learned critic.
§3–§4.3; standard RL-for-LLM assumption imported from GRPO literature.
domain assumption Equal-weight next-day open-to-open 35% long–short portfolios with Rf=0 and no frictions are an adequate economic test of sentiment signal quality.
§5 portfolio construction and evaluation; all headline metrics depend on it.
domain assumption Logit-to-score conversion from FinDPO yields continuous scores comparable across causal LLM sentiment systems.
§5; required for portfolio ranking of FinSMART vs FinDPO.

invented entities (2)

dual-filter asymmetric trading reward R_trade(d, y) no independent evidence
purpose: Map predicted sentiment and market state into a scalar RL reward that requires both directional correctness and economically meaningful alpha.
Core paper-specific objective in §4.2; not a physical entity but a new stipulated reward object on which all learning depends. No independent evidence outside the backtest that these magnitudes are uniquely correct.
FinSMART market-aligned RL training pipeline (NER + sentiment gate + GRPO on market outcomes) no independent evidence
purpose: End-to-end framework claimed as first market-aligned RL sentiment system with continual retrainability.
Composite invented system of the paper; falsifiable only via external replication of the trading results.

pith-pipeline@v1.2.0-daily-grok45 · 17473 in / 4463 out tokens · 97681 ms · 2026-07-31T17:00:15.503708+00:00 · methodology

0 comments

read the original abstract

Recent advances in Generative AI have substantially improved financial sentiment analysis through post-trained financial large language models (LLMs). However, existing approaches remain confined to a market-agnostic, supervised learning paradigm that relies on limited, static and human-annotated datasets, and thus are incapable of adapting to evolving market conditions. To address this limitation, we introduce FinSMART, the first market-aligned reinforcement learning framework for financial sentiment analysis, which directly optimizes sentiment signals using realized market outcomes. To deal with the noisy, non-stationary, and multifactorial nature of financial markets, FinSMART incorporates a signal extraction pipeline that combines market-aware data filtering with a discrete asymmetric trading reward, enabling stable reinforcement learning from economically meaningful market feedback. Experimental results demonstrate that FinSMART significantly outperforms existing state-of-the-art methods in profitability, risk-adjusted performance, and sentiment signal quality, improving cumulative trading returns by 220% over the strongest baseline. Uniquely, the FinSMART framework naturally supports market-aware retraining, at any point in time, by replacing costly manual annotation with newly observed financial articles and their realized market outcomes. Such a retraining strategy enables the model to continuously adapt to changing market dynamics, resulting in consistent performance gains over its static counterpart. These findings demonstrate the practical applicability of market-aligned reinforcement learning and highlight its potential as a next-generation paradigm for developing adaptive financial LLMs.

Figures

Figures reproduced from arXiv: 2607.28127 by Danilo Mandic, Giorgos Iacovides, Wuyang Zhou.

**Figure 2.** Figure 2: The FinSMART framework for market-aligned reinforcement learning. FinSMART transforms realized market [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** Figure 3: Proposed framework for our sentiment-driven port [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

**Figure 4.** Figure 4: Cumulative returns of sentiment-based long-short portfolios and the impact of market-aligned retraining. Left: Out [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

discussion (0)

Reference graph

Works this paper leans on

20 extracted references · 5 linked inside Pith

[1]

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

Pith/arXiv arXiv 2019
[2]

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. InThe Second Conference on Parsimony and Learning (Recent Spotlight Track)

2025
[3]

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. InProceedings of the Fourth ACM International Conference on AI in Finance. 357–364

2023
[4]

Aaron Grattafiori et al . 2024. The llama 3 herd of models.arXiv preprint arXiv:2407.21783(2024)

Pith/arXiv arXiv 2024
[5]

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. InInternational Conference on Learning Representations. https://openreview.net/forum?id=nZeVKeeFYf9

2022
[6]

Clayton Hutto and Eric Gilbert. 2014. VADER: A parsimonious rule-based model for sentiment analysis of social media text. InProceedings of the International AAAI Conference on Web and Social Media, Vol. 8. 216–225

2014
[7]

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

2024
[9]

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

2019
[10]

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

2021
[11]

Tim Loughran and Bill Mcdonald. 2011. When Is a Liability NOT a Liability? Textual Analysis, Dictionaries, and 10-Ks.The Journal of Finance66, 35 – 65

2011
[12]

Yangtuo Peng and Hui Jiang. 2016. Leverage Financial News to Predict Stock Price Movements Using Word Embeddings and Deep Neural Networks. InPro- ceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. Association for Computational Linguistics, 374–379

2016
[13]

John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov
[14]

Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y.K. Li, Y. Wu, and Daya Guo. 2024. DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models. https: //arxiv.org/abs/2402.03300

Pith/arXiv arXiv 2024
[15]

Stone, Dexter C

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

1966
[16]

Dimitris Vamvourellis and Dhagash Mehta. 2025. Reasoning or Overthink- ing: Evaluating Large Language Models on Financial Sentiment Analysis. In Proceedings of the 6th ACM International Conference on AI in Finance (ICAIF ’25). Association for Computing Machinery, New York, NY, USA, 299–307. doi:10.1145/3768292.3770341

arXiv 2025
[17]

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

2023
[18]

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

2023
[19]

Boyu Zhang, Hongyang Yang, and Xiao-Yang Liu. 2023. Instruct-FinGPT: Fi- nancial Sentiment Analysis by Instruction Tuning of General-Purpose Large Language Models.ArXivabs/2306.12659 (2023)

Pith/arXiv arXiv 2023
[20]

Wenbin Zhang and Steven Skiena. 2010. Trading Strategies to Exploit Blog and News Sentiment.Proceedings of the International AAAI Conference on Web and Social Media4, 1 (May 2010), 375–378

2010
[2017]

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

Pith/arXiv arXiv 2017

[1] [1]

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

Pith/arXiv arXiv 2019

[2] [2]

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. InThe Second Conference on Parsimony and Learning (Recent Spotlight Track)

2025

[3] [3]

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. InProceedings of the Fourth ACM International Conference on AI in Finance. 357–364

2023

[4] [4]

Aaron Grattafiori et al . 2024. The llama 3 herd of models.arXiv preprint arXiv:2407.21783(2024)

Pith/arXiv arXiv 2024

[5] [5]

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. InInternational Conference on Learning Representations. https://openreview.net/forum?id=nZeVKeeFYf9

2022

[6] [6]

Clayton Hutto and Eric Gilbert. 2014. VADER: A parsimonious rule-based model for sentiment analysis of social media text. InProceedings of the International AAAI Conference on Web and Social Media, Vol. 8. 216–225

2014

[7] [7]

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

2024

[8] [9]

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

2019

[9] [10]

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

2021

[10] [11]

Tim Loughran and Bill Mcdonald. 2011. When Is a Liability NOT a Liability? Textual Analysis, Dictionaries, and 10-Ks.The Journal of Finance66, 35 – 65

2011

[11] [12]

Yangtuo Peng and Hui Jiang. 2016. Leverage Financial News to Predict Stock Price Movements Using Word Embeddings and Deep Neural Networks. InPro- ceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. Association for Computational Linguistics, 374–379

2016

[12] [13]

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

[13] [14]

Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y.K. Li, Y. Wu, and Daya Guo. 2024. DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models. https: //arxiv.org/abs/2402.03300

Pith/arXiv arXiv 2024

[14] [15]

Stone, Dexter C

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

1966

[15] [16]

Dimitris Vamvourellis and Dhagash Mehta. 2025. Reasoning or Overthink- ing: Evaluating Large Language Models on Financial Sentiment Analysis. In Proceedings of the 6th ACM International Conference on AI in Finance (ICAIF ’25). Association for Computing Machinery, New York, NY, USA, 299–307. doi:10.1145/3768292.3770341

arXiv 2025

[16] [17]

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

2023

[17] [18]

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

2023

[18] [19]

Boyu Zhang, Hongyang Yang, and Xiao-Yang Liu. 2023. Instruct-FinGPT: Fi- nancial Sentiment Analysis by Instruction Tuning of General-Purpose Large Language Models.ArXivabs/2306.12659 (2023)

Pith/arXiv arXiv 2023

[19] [20]

Wenbin Zhang and Steven Skiena. 2010. Trading Strategies to Exploit Blog and News Sentiment.Proceedings of the International AAAI Conference on Web and Social Media4, 1 (May 2010), 375–378

2010

[20] [2017]

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

Pith/arXiv arXiv 2017