From Factor Models to Deep Learning: Machine Learning in Reshaping Empirical Asset Pricing
From Factor Models to Deep Learning: Machine Learning in Reshaping Empirical Asset Pricing
This paper provides a comprehensive review of machine learning (ML) applications in empirical asset pricing, covering the evolution from traditional factor models to advanced deep learning techniques. It systematically categorizes ML approaches—supervised, unsupervised, semi-supervised, and reinforcement learning—and demonstrates how they address limitations of classical models like CAPM and Fama-French, such as linearity assumptions and poor handling of high-dimensional data. Key methods include temporal models (RNN, LSTM, N-BEATS, TS-Mixer, Transformers) for time-series forecasting, and reinforcement learning for dynamic portfolio optimization. The review highlights significant improvements in predictive accuracy and the ability to incorporate non-traditional data sources (text, images). However, it also critically discusses challenges like overfitting, lack of explainability, and data quality issues, identifying future research directions. This survey is unique in its focus on ML algorithm development for asset pricing, offering a valuable interdisciplinary resource for researchers and practitioners.
Highlights
- 1Comprehensive review of ML applications in empirical asset pricing, bridging finance and computer science perspectives.
- 2Detailed taxonomy of ML models (supervised, unsupervised, semi-supervised, RL) for return prediction, portfolio optimization, and risk assessment.
- 3Critical evaluation of challenges: explainability, overfitting, and integration of heterogeneous data sources.
- 4First survey emphasizing ML algorithm development specifically for asset pricing.
Methods
- MTraditional factor models (CAPM, Fama-French, PCA, IPCA) for risk premia estimation.
- MSupervised learning (linear regression, decision trees, random forests, gradient boosting, deep neural networks) for return prediction.
- MTemporal models (RNN, LSTM, N-BEATS, TS-Mixer, Transformer) for time-series forecasting.
- MReinforcement learning for portfolio optimization.
Results
- RML models outperform traditional factor models in predictive accuracy by capturing nonlinearities and high-dimensional interactions.
- RDeep learning architectures (e.g., TS-Mixer, Transformer) achieve state-of-the-art results in stock return and movement prediction.
- RIntegration of heterogeneous data (text, images) enhances asset pricing models.
- RChallenges remain in model interpretability, overfitting, and data quality, requiring further research.
Analyze Paper
Generate insights from "From Factor Models to Deep Learning: Machine Learning in Res...".