The List / Trading, Investment, and Portfolio Management / paper
Exploring the Synergy of Quantitative Factors and Newsflow Representations from Large Language Models for Stock Return Prediction
Abstract
In quantitative investing, return prediction supports various tasks, including stock selection, portfolio optimization, and risk management. Quantitative factors, such as valuation, quality, and growth, capture various characteristics of stocks. Unstructured data, like news and transcripts, has attracted growing attention, driven by recent advances in large language models (LLMs). This paper examines effective methods for leveraging multimodal factors and newsflow in return prediction and stock selection. First, we introduce a fusion learning framework to learn a unified representation from factors and newsflow representations generated by an LLM. Within this framework, we compare three methods of different architectural complexities: representation combination, representation summation, and attentive representations. Next, building on the limitation of fusion learning observed in empirical comparison, we explore the mixture model that adaptively combines predictions made by single modalities and their fusion. To mitigate the training instability of the mixture model, we introduce a decoupled training approach with theoretical insights. Finally, our experiments on real investment universes yield several insights into effective multimodal modeling of factors and news for stock return prediction and selection.
In one sentence
Summary
This research addresses the challenge of effectively integrating structured quantitative factors with unstructured financial newsflow for stock return prediction. The authors propose a multimodal fusion learning framework comparing representation combination, summation, and attentive methods, alongside a novel mixture model that adaptively weights single-modality and fused predictions. To overcome training instability in mixture models caused by entangled gradient variance, they introduce a decoupled training approach that independently trains prediction components and aligns mixture probabilities with their relative performance via KL divergence minimization. This method theoretically ensures that the mixture can outperform individual components by leveraging their complementary strengths. Experiments are conducted on three investment universes: North American, Emerging Markets, and European stocks, using commercial news data and standard quantitative factors. The study evaluates long-only and long-short portfolios rebalanced monthly, comparing performance against baselines like factors-only, news-only, and other fusion methods. Results show that while simple fusion (concatenation) works well in efficient markets like North America, it can dilute factor signals in less efficient markets like Emerging Markets. The decoupled mixture model demonstrates robustness across all universes, achieving the highest annualized returns and Sharpe ratios in long-short portfolios, particularly in the North American and Emerging Markets datasets. The paper also investigates the impact of fine-tuning the LLM (DeBERTa) via LoRA, finding that it does not consistently improve performance, suggesting that pre-trained representations may suffice for this task. Key findings indicate that the predictive relevance of news varies significantly across markets and time, often overlapping with factor information. The mixture model’s ability to adaptively downweight noisy or redundant news when factors are strong makes it superior for alpha mining in volatile or less efficient environments. Limitations include the dependency on the quality of news data and the potential for overfitting if the mixture weights are not properly regularized. The work provides practical insights for quantitative investors on when to use fusion versus mixture approaches, highlighting that simple concatenation is effective in stable regimes, while adaptive mixing is crucial for robustness across diverse market conditions.
Research problem
The paper addresses the problem of effectively combining structured quantitative factors with unstructured financial newsflow representations generated by LLMs for stock return prediction. It specifically investigates how to mitigate the dilution of factor signals when news is irrelevant or redundant, and how to stabilize the training of mixture models that adaptively combine these modalities.
Core contributions
- A multimodal fusion learning framework comparing representation combination, summation, and attentive fusion methods.
- A novel mixture model that adaptively weights predictions from single modalities and their fusion.
- A decoupled training method with theoretical insights to mitigate gradient variance instability in mixture models.
- Empirical analysis across three investment universes demonstrating the varying predictive relevance of news vs. factors.
- Insights into the impact of LLM fine-tuning on multimodal financial prediction tasks.
Method
- LLM (DeBERTa) generates newsflow representations aggregated into vectors.
- Fusion learning combines factors and news via concatenation, summation, or attention.
- Mixture model uses a gating network to weight factors-based and fusion-based predictions.
- Decoupled training minimizes independent squared errors and KL divergence for probability alignment.
- LoRA fine-tuning applied to DeBERTa in ablation studies.
- Backtesting with long-only (top decile) and long-short (top/bottom decile) portfolios.
Results & findings
- Fusion Combination outperforms other fusion methods in North American long-only portfolios.
- Mixture Decoupled achieves the highest Sharpe ratio (1.78) and annualized return (33.77%) in North American long-short portfolios.
- In Emerging Markets, Mixture Decoupled matches Factors Alone in long-short but outperforms in long-only.
- LLM fine-tuning via LoRA does not consistently improve performance across universes.
- News relevance varies by market; fusion can dilute factor signals in less efficient markets.
- Decoupled training stabilizes convergence and improves overall mixture model performance.
Data & experiments
- Three investment universes: North American, Emerging Markets, and European stocks.
- Commercial news data and standard quantitative factors (valuation, quality, growth, etc.).
- Long-only and long-short portfolio backtests with monthly rebalancing.
- Comparison against baselines: factors alone, news alone, FININ, and various fusion methods.
- Evaluation metrics: annualized return, Sharpe ratio, MAPE, and Information Coefficient (IC).
Limitations & caveats
- Performance of fusion methods is universe-dependent; news may be redundant in efficient markets.
- LLM fine-tuning benefits are inconsistent, suggesting pre-trained models may be sufficient.
- Mixture model complexity may increase computational cost compared to simple fusion.
- Results are based on historical data; future market regime changes may affect robustness.
- News data quality and coverage may vary across different commercial vendors.
Curator notes
Discovery status: round2_promoted. Source links: A Multimodal Foundation Agent for Financial Trading: Tool-Augmented, Diversified, and Generalist; AlphaFin: Benchmarking Financial Analysis with Retrieval-Augmented Stock-Chain Framework; FinBERT: Financial Sentiment Analysis with Pre-trained Language Models; FinTral: A Family of GPT-4 Level Multimodal Financial Large Language Models; Mixing It Up: The Cocktail Effect of Multi-Task Fine-Tuning on LLM Performance - A Case Study in Finance; Open-FinLLMs: Open Multimodal Large Language Models for F