The List / Trading, Investment, and Portfolio Management / paper
Fine-Tuning Large Language Models for Stock Return Prediction Using Newsflow
Abstract
Large language models (LLMs) and their fine-tuning techniques have demonstrated superior performance in various language understanding and generation tasks.This paper explores fine-tuning LLMs for predicting stock returns with financial newsflow.Return prediction is fundamental for subsequent tasks like portfolio construction and optimization in quantitative investing. We formulate the model to include a text representation and forecasting modules. We propose to compare the encoder-only and decoder-only LLMs, considering they generate text representations in distinct ways.The impact of these different representations on return forecasting remains an open question.Meanwhile, we compare two simple methods of integrating LLMs’ token-level representations into the forecasting module.The experiments on real investment universes reveal that:(1) aggregated representations from LLMs’ token-level embeddings generally produce return predictions that enhance the performance of long-only and long-short portfolios;(2) in the relatively large investment universe, the decoder LLMs-based prediction model leads to stronger portfolios, whereas in the small universes, there are no consistent winners;(3) return predictions derived from LLMs’ text representations are a strong signal for portfolio construction, outperforming conventional sentiment scores.These findings shed light on developing suitable LLM fine-tuning methods for return prediction-based portfolio construction.
In one sentence
Summary
The paper addresses the challenge of extracting predictive signals from financial news for quantitative stock picking by proposing a direct news-to-return forecasting framework. Unlike traditional workflows that rely on intermediate feature extraction such as sentiment scoring, this approach fine-tunes Large Language Models (LLMs) to map text sequences directly to forward stock returns. The authors investigate the impact of model architecture by comparing encoder-only (DeBERTa) and decoder-only (Mistral, Llama3) LLMs, and evaluate two representation integration strategies: bottleneck representations using an end-of-sequence token and aggregated representations via token averaging. This design aims to determine which text encoding methods best capture semantic information relevant to future price movements. Experiments were conducted on real-world financial news data from 2003 to 2019 across North American, European, and Emerging Market investment universes. The models were fine-tuned using Low-Rank Adaptation (LoRA) to minimize mean squared error between predicted and actual monthly forward returns. The evaluation included decile-wise metrics such as RMSE, precision, and return, alongside backtesting of long-only and long-short portfolios constructed from the top and bottom deciles of predictions. The study compared these LLM-based portfolios against benchmarks using FinBERT and FinVader sentiment scores, ensuring a rigorous assessment of predictive power and economic value in out-of-sample testing periods. Results indicate that aggregated representations generally enhance portfolio performance, with decoder-based models like Mistral leading in larger universes. The LLM-derived return predictions significantly outperformed conventional sentiment-based portfolios in both annualized returns and Sharpe ratios, particularly in long-short strategies where the short side mitigated volatility. However, the study notes limitations, including inconsistent performance of encoder-only models in large universes and varying results for Llama across different markets. These findings suggest that while LLMs offer strong alpha signals, the choice of architecture and representation method must be tailored to the specific investment universe and strategy, highlighting the need for further exploration of model scaling and robustness.
Research problem
The paper tackles the problem of effectively utilizing financial newsflow for stock return prediction in quantitative investing. It seeks to determine whether fine-tuning LLMs for direct return forecasting outperforms traditional feature-extraction methods like sentiment analysis. Specifically, it investigates how different LLM architectures (encoder vs. decoder) and representation aggregation techniques impact the quality of return signals for portfolio construction.
Core contributions
- Proposes a direct news-to-return forecasting framework that bypasses intermediate feature extraction.
- Compares encoder-only (DeBERTa) and decoder-only (Mistral, Llama3) LLMs for financial text representation.
- Evaluates bottleneck versus aggregated token representations for integrating LLM embeddings into forecasting modules.
- Demonstrates that LLM-based return predictions outperform conventional sentiment-based portfolios in backtests.
- Provides empirical evidence on the robustness of different LLM architectures across diverse global investment universes.
Method
- Concatenates financial news within a look-back window into a single text sequence.
- Fine-tunes pre-trained LLMs using Low-Rank Adaptation (LoRA) with rank 4.
- Uses a dense layer as the forecasting module to map text embeddings to returns.
- Compares encoder-only (DeBERTa) and decoder-only (Mistral, Llama3) architectures.
- Implements bottleneck representation by appending an EOS token and using its embedding.
- Implements aggregated representation by averaging all token embeddings.
- Trains models to minimize mean squared error between predicted and actual forward returns.
- Constructs portfolios by ranking stocks based on predicted returns and selecting top/bottom deciles.
Results & findings
- Aggregated representations generally produce better return predictions than bottleneck representations.
- Decoder-only LLMs (Mistral) lead in larger investment universes, while results are mixed in smaller ones.
- LLM-based portfolios outperform sentiment-based portfolios in both long-only and long-short strategies.
- Long-short portfolios using LLM predictions show smoother return curves and higher Sharpe ratios.
- Mistral demonstrates the most robust performance across different market universes.
- Encoder-only DeBERTa underperforms in large universes, possibly due to model size or architecture.
- Decile precision is a more indicative metric for portfolio performance than raw RMSE in extreme deciles.
Data & experiments
- Financial news data from 2003 to 2019 linked to company identifiers.
- Three investment universes: North America (630 stocks), Europe (350 stocks), and Emerging Markets (370 stocks).
- Monthly forward returns used as the target variable for regression.
- Backtesting of long-only and long-short portfolios based on decile rankings of predictions.
- Comparison against FinBERT and FinVader sentiment-based baselines.
- Evaluation metrics include decile RMSE, precision, annualized returns, and Sharpe ratios.
Limitations & caveats
- Performance of encoder-only models varies significantly across different universe sizes.
- Llama's performance is inconsistent compared to Mistral, requiring further exploration.
- The study focuses on monthly returns, which may not capture high-frequency trading dynamics.
- Results are based on historical data (2003-2019), which may not reflect current market conditions.
- The impact of model size on performance is not fully disentangled from architectural differences.
- Sentiment baselines may not represent the state-of-the-art in feature extraction methods.
Curator notes
Discovery status: round3_promoted. Source links: Exploring the Synergy of Quantitative Factors and Newsflow Representations from Large Language Models for Stock Return Prediction; Large Language Models in equity markets: applications, techniques, and insights