Paper: arXiv 2507.18417
Authors: Giorgos Iacovides, Wuyang Zhou, Danilo Mandic
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).
Complexity vs Empirical Score
- Math Complexity: 3.0/10
- Empirical Rigor: 8.5/10
- Quadrant: Street Traders — empirical and applied, lower math complexity
Why this score: The paper’s primary innovation is a novel framework for training and evaluating a Large Language Model, focusing on practical implementation details and backtest results with specific financial metrics, rather than complex mathematical derivations.
Research Flowchart
flowchart TD
A["Research Goal: Develop robust FinSA model for trading"] --> B["FinDPO Framework<br>LLM Pref. Opt. via DPO"]
B --> C["Data: Financial Sentiment Benchmarks"]
C --> D["Compute: Logit-to-Score Conversion<br>→ Continuous Probabilities"]
D --> E["Simulation: Portfolio Strategy<br>with Trans. Costs"]
E --> F{"Outcomes"}
F --> G["SOTA Sentiment Accuracy<br>+11% over SFT"]
F --> H["Realized Portfolio Returns<br>67% Annual / Sharpe 2.0"]
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper proposes FinDPO, a finance‑specific LLM framework aligned with human preferences via Direct Preference Optimization for financial sentiment analysis and trading.
Source sentence
We introduce FinDPO, the first finance-specific LLM framework based on post-training human preference alignment via Direct Preference Optimization (DPO).
Assumes 5
- Transaction costs are assumed to lie between 1 and 5 basis points, typical for U.S. equities. execution
Source sentence
we varied the transaction cost parameter 𝑘from 1 to 5 basis points (bps), which aligns with the typical 2–5 bps range observed in U.S. equities.
- The investable universe consists of liquid S&P 500 constituent stocks. market
Source sentence
Given that our investable universe consisted of relatively liquid instruments, specifically the S&P 500 constituents
- A daily rebalancing, long‑short strategy with equal long and short position sizes is used. execution
Source sentence
Given that our investable universe consisted of relatively liquid instruments, specifically the S&P 500 constituents, we varied the transaction cost parameter…
- Risk‑free rate is set to 0% for performance calculations. statistical
Source sentence
a 0% rate is commonly used and was adopted in our analysis.
Show 1 more
- Preference pairs are generated by prompting a frozen reference LLM to produce dispreferred responses. model
Source sentence
we prompted the reference model, 𝜋ref, with ˜𝑥𝑖 to generate sentiment predictions.
Data
2015 – 2021 · S&P 500 stocks · equities · daily
- Financial PhraseBank (FPB) Dataset text-news public
Source sentence
The FPB dataset consists of 4,840 samples which were randomly extracted from financial news articles.
- Twitter Financial News Sentiment (TFNS) Dataset text-news public
Source sentence
The TFNS dataset includes 11,930 tweets with content from the financial domain.
- GPT‑labelled Financial News (NWGI) Dataset text-news public
Source sentence
The NWGI dataset consists of 16,200 financial news articles.
- S&P 500 price data market-prices public
Source sentence
Given that our investable universe consisted of relatively liquid instruments, specifically the S&P 500 constituents
Costs
Unclear - sources disagree
Source sentence (the authors' words)
we varied the transaction cost parameter 𝑘from 1 to 5 basis points (bps), which aligns with the typical 2–5 bps range observed in U.S. equities.
Finds 4
Sharpe 2.00 (benchmark 0.62) · annual return 67.0% · total return 747.1%
- FinDPO achieves state‑of‑the‑art sentiment classification performance, improving average weighted F1 by 11% over the SOTA FinGPT v3.3 model. predictive-accuracy
Source sentence
our proposed FinDPO model has achieved a weighted F1 score that outperforms the SOTA FinGPT v3.3 model by 11% on the average.
- In a frictionless setting FinDPO delivers cumulative return of 747.10% (annualized 111.78%) with Sharpe 3.41, Sortino 6.05 and Calmar 11.94, far surpassing all baselines. performance
Source sentence
FinDPO (Ours) 747.10 111.78 3.41 6.05 11.94
- FinDPO outperforms the prior best sentiment‑based portfolio (FinLlama) by 187% in cumulative returns and improves Sharpe ratio by 1.1. performance
Source sentence
... achieving a 187% improvement in cumulative returns.
- Even under realistic transaction costs of up to 5 bps, FinDPO maintains substantial positive returns and strong risk‑adjusted performance. performance
Source sentence
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).
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 40 related papers). Coverage is partial, so this is neither replication nor refutation.
1 paper in the archive : Trade-R1: Bridging Verifiable Rewards to Stochastic …
Code
No repository by the authors found in the paper.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.
Cited References (37)
- Good debt or bad debt: Detecting semantic orientations in economic texts (2014)
- Twitter Financial News Sentiment (2022) — link
- News with GPT instructions (2023) — link
- Fixing weight decay regularization in adam (2017) — link
- Lo RA : Low-Rank Adaptation of Large Language Models (2022)
- Direct Preference Optimization: Your Language Model is Secretly a Reward Model (2023)
- The llama 3 herd of models (2024) — link
- The stock sonar—sentiment analysis of stocks based on a hybrid approach (2011)
and 29 more
- Which news moves stock prices? A textual analysis (2013)
- BERT-base-NER (2021)
- When Is a Liability NOT a Liability? Textual Analysis, Dictionaries, and 10- Ks (2011)
- The General Inquirer: A Computer Approach to Content Analysis (1966)
- Vader: A parsimonious rule-based model for sentiment analysis of social media text (2014)
- FinBERT : Financial sentiment analysis with pre-trained language models (2019) — link
- FinLlama: LLM-Based Financial Sentiment Analysis for Algorithmic Trading (2024)
- FinGPT : Instruction Tuning Benchmark for Open-Source Large Language Models in Financial Datasets — → 2310.04793
- Taming Overconfidence in LLM s: Reward Calibration in RLHF (2025)
- On Calibration of Modern Neural Networks (2017)
- Treasury yield 10 years historical data. (2023) — link
- Training language models to follow instructions with human feedback (2022)
- Leverage Financial News to Predict Stock Price Movements Using Word Embeddings and Deep Neural Networks (2016)
- Trading Strategies to Exploit Blog and News Sentiment (2010)
- Predicting Returns With Text Data (2019) — link
- Instruct- FinGPT : Financial Sentiment Analysis by Instruction Tuning of General-Purpose Large Language Models — → 2306.12659
- Finbert-fomc: Fine-tuned finbert model with sentiment focus method for enhancing sentiment analysis of fomc minutes (2023)
- Simple Statistical Gradient-Following Algorithms for Connectionist Reinforcement Learning (2004)
- Proximal policy optimization algorithms (2017) — link
- Is Reinforcement Learning (Not) for Natural Language Processing: Benchmarks, Baselines, and Building Blocks for Natural Language Policy Optimization (2023)
- SFT Memorizes, RL Generalizes: A Comparative Study of Foundation Model Post-training (2025)
- Scilab — link
- Scilab — link
- The Computational Geometry Algorithms Library — link
- The Computational Geometry Algorithms Library — link
- The Parmap library — link
- The Parmap library — link
- BabyJ : from Object Based to Class Based Programming via Types (2003)
- tag-sound
2 of 37 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.