Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Exploring the Synergy of Quantitative Factors and Newsflow Representations from Large Language Models for Stock Return Prediction

Tian Guo, Emmanuel Hauptmann

TL;DR

This work addresses stock return prediction by fusing quantitative factors with unstructured newsflow representations generated by large language models. It introduces a fusion learning framework that creates unified representations from factor and news modalities, and compares three fusion architectures, followed by a mixture model that adaptively combines factors-based and fusion-based predictions. To stabilize the training of the mixture, the authors propose a decoupled training scheme grounded in distribution matching via KL divergence between a trainable and a target mixture distribution, with theoretical support. Experiments across North American, Emerging Markets, and European universes show that simple fusion can outperform single-modality baselines, while the decoupled mixture yields robust performance across markets; fine-tuning the LLM has mixed effects depending on market efficiency. Overall, the paper provides practical guidance for multimodal modeling in finance, highlighting when to rely on fusion versus mixture methods and how to train them for stability and cross-market robustness.

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.

Exploring the Synergy of Quantitative Factors and Newsflow Representations from Large Language Models for Stock Return Prediction

TL;DR

This work addresses stock return prediction by fusing quantitative factors with unstructured newsflow representations generated by large language models. It introduces a fusion learning framework that creates unified representations from factor and news modalities, and compares three fusion architectures, followed by a mixture model that adaptively combines factors-based and fusion-based predictions. To stabilize the training of the mixture, the authors propose a decoupled training scheme grounded in distribution matching via KL divergence between a trainable and a target mixture distribution, with theoretical support. Experiments across North American, Emerging Markets, and European universes show that simple fusion can outperform single-modality baselines, while the decoupled mixture yields robust performance across markets; fine-tuning the LLM has mixed effects depending on market efficiency. Overall, the paper provides practical guidance for multimodal modeling in finance, highlighting when to rely on fusion versus mixture methods and how to train them for stability and cross-market robustness.

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.
Paper Structure (23 sections, 2 theorems, 34 equations, 16 figures, 11 tables)

This paper contains 23 sections, 2 theorems, 34 equations, 16 figures, 11 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Factors and Newsflow for Return Prediction and Stock Selection
  4. Problem Statement
  5. Methodologies
  6. Fusion Learning over Factors and Newsflow Representations
  7. Mixture Modeling over Factors-based and Fusion-based Predictions
  8. Experiments
  9. Conclusion
  10. Future Work
  11. Appendix
  12. Proof of Proposition 1 and Discussion
  13. Proof of Proposition 2 and Discussion
  14. Theoretical Insights of Decoupled Training
  15. Experiment Details and Full Results
  16. ...and 8 more sections

Key Result

Proposition 1

Consider stochastic gradient descent with instances sampled as $\{ \mathbf{x}_f, \mathbf{x}_n, r \} \sim \mathcal{D}$. Let $\hat{r}$ denote the model's prediction. Let $i$ index a prediction component in the mixture model, and define the gradient signal for component $i$ as $\zeta_i:= \left( r - \ha By contrast, under the standalone training of component $i$, the gradient variance is:

Figures (16)

  • Figure 1: Workflow using quantitative factors and newsflow for return prediction and stock selection.
  • Figure 2: Illustration of fusion learning and mixture model.
  • Figure 3: Illustrative comparison of different methods' portfolio performance (North American Universe).
  • Figure 4: (a) Training error curves of the mixture model's prediction components by different training methods. (b) Illustration of the alignment between each component’s prediction errors (top panel) and mixture probabilities (bottom panel) learned via the distribution matching in decoupled training. Note that the x-axis represents test samples ordered arbitrarily (not a time series).
  • Figure 5: Performance Charts.
  • ...and 11 more figures

Theorems & Definitions (4)

  • Proposition 1: Entangled Gradient Variance
  • Proposition 2: Optimal Mixture Weights and the Relation to Eq. \ref{['eq:target_distribution']}
  • proof
  • proof