TLDR: This paper explores combining traditional quantitative stock factors with news insights from Large Language Models (LLMs) for better stock return prediction. It introduces a “fusion learning” framework, finding simple representation combination effective. It also proposes a “mixture model” with “decoupled training” to adaptively combine factor-based and fusion-based predictions, overcoming instability and improving performance. Experiments show the mixture model is robust, and LLM fine-tuning’s impact varies by market efficiency.
Predicting stock returns is a cornerstone of quantitative investing, influencing decisions from selecting individual stocks to managing overall portfolio risk. Traditionally, investors have relied on “quantitative factors” – numerical characteristics of stocks like their valuation, quality, and growth prospects. These factors are derived from structured data such as prices, economic indicators, and analyst estimates, and have a long history of demonstrating predictive power in financial markets.
However, the financial world is also awash with unstructured data, particularly financial news. With the rapid advancements in large language models (LLMs), there’s a growing interest in how this textual information can be harnessed for better investment decisions. A recent research paper delves into this very area, examining how to effectively combine these traditional quantitative factors with insights extracted from financial news using LLMs to predict stock returns and guide stock selection. You can read the full paper here: Exploring the Synergy of Quantitative Factors and Newsflow Representations from Large Language Models for Stock Return Prediction.
The Challenge of Combining Different Data Types
The core challenge lies in integrating two fundamentally different types of data: structured numerical factors and unstructured textual news. Quantitative factors are stable and grounded in financial theory, offering consistent predictive power. News, on the other hand, is often noisy, context-dependent, and its relevance can fluctuate significantly. Sometimes, news might offer little new information beyond what factors already capture, potentially diluting the predictive signal if not handled carefully.
Fusion Learning: Blending Information
The researchers first explored a “fusion learning” framework. This approach aims to create a single, unified representation by combining quantitative factors with “newsflow representations” – numerical vectors generated by an LLM that capture the essence of financial news. Within this framework, they tested three main methods:
- Representation Combination: This is the simplest method, treating each dimension of the news representation as an additional feature alongside the numerical factors. These are then concatenated and processed by a simple neural network layer.
- Representation Summation: This method projects both factors and news representations into a shared space and then sums them, encouraging alignment between the two modalities.
- Attentive Representation: An extension of summation, this method introduces adaptive weights to the factor and news representations, allowing the model to dynamically adjust the importance of each modality based on the specific instance.
Surprisingly, experiments showed that the relatively simple “representation combination” method often outperformed the more complex fusion techniques. This suggests that in the often-noisy financial environment, straightforward model architectures that work across different modality-specific representations can be highly effective.
Mixture Models: Adaptive Prediction
While fusion learning showed promise, it had a limitation: it sometimes struggled when news was less relevant, potentially diluting the strong signal from factors. To address this, the researchers developed a “mixture model.” This model doesn’t just fuse representations; it adaptively combines predictions made separately by a factors-based component and a fusion-based component. This allows the model to leverage the strengths of each component, adapting to situations where factors might be more dominant or where news provides crucial incremental information.
A significant hurdle with conventional training of such mixture models was instability. The researchers identified that this instability stemmed from “entangled gradient variance” – essentially, the updates for one part of the model were negatively affected by the noise and interactions with other parts. To overcome this, they introduced a “decoupled training” approach. This method trains each prediction component independently to maximize its individual predictive capacity, while simultaneously learning how to optimally combine their predictions based on their actual performance. This decoupled approach led to more stable training and significantly improved overall performance.
Also Read:
- Interpreting Digital Emotions: A Hybrid Framework for Nuanced Sentiment Analysis
- Evaluating Language Models on Real-World Uncertainty with OPENESTIMATE
Key Findings from Real-World Experiments
The research involved extensive experiments using real investment data across North American (NA), European (EU), and Emerging Markets (EM) universes. The findings provided several valuable insights:
- Fusion Learning’s Effectiveness: Integrating factors and newsflow can indeed create powerful predictive representations, with the simple “representation combination” method often leading the pack.
- Mixture Model’s Adaptability: The mixture model, especially with decoupled training, achieved comparable or superior performance to fusion learning. Its ability to adaptively weigh predictions from different sources proved particularly beneficial in markets where the relevance of news versus factors can vary. The decoupled training was crucial for unlocking this potential.
- Inconsistent Impact of LLM Fine-Tuning: One intriguing finding was that fine-tuning the LLM (DeBERTa in this case) during the training of multimodal prediction models did not consistently improve performance. Its impact varied significantly across different investment universes. In highly efficient markets like North America, fine-tuning sometimes led to overfitting noisy or already-priced news information, potentially degrading performance. In less efficient, more heterogeneous markets like EM and EU, fine-tuning could lead to improvements, suggesting that news in these markets might contain more unpriced, nuanced information that benefits from specialized LLM adaptation.
This research offers valuable insights into building more robust and adaptive stock return prediction models by strategically combining traditional quantitative factors with the rich, yet challenging, information from financial news, all powered by advanced large language models. The findings underscore the importance of not just integrating data, but doing so with sophisticated model designs and training methodologies tailored to the unique characteristics of financial markets.
