time：2024-04-01
Esfandiar Maasoumi, Jianqiu Wang, Zhuo Wang, Ke Wu
This paper focuses on identifying risk factors that significantly explain stock returns, introducing a novel method called Automatic Debiased Machine Learning (ADML). This method helps to estimate the partial pricing effect of individual factors while controlling for a large number of confounding factors under a nonlinear stochastic discount factor (SDF) assumption. The ADML method overcomes issues commonly found in traditional machine learning techniques, such as biased estimation, non-robustness, and overfitting. The findings suggest that the factors identified through ADML outperform the Fama–French sparse factors and factors derived from the double-selection LASSO method when using a linear factor model.
The study identifies between 30 and 50 factors that exhibit significant but declining pricing power in explaining stock returns. The simulation results show that ADML accurately identifies important factors, providing robust estimates with asymptotic normality. Additionally, the paper conducts a classical spanning test, showing that ADML-based models surpass traditional models like the Fama–French models in terms of the number of factors explained and the magnitude of absolute factor alphas. The ADML models also demonstrate better explanatory power, especially for intangible factors.
The paper further explores the time-series variation in the pricing power of each factor and finds that, while some well-known factors such as SMB, HML, and ROE maintain persistent pricing power, the overall significance of factors declines over time. The ADML approach identifies a much larger set of significant factors compared to linear models, such as OLS or the double-selection LASSO method, indicating the importance of nonlinear pricing information.
Finally, the study performs robustness checks using the Chinese stock market, which is influenced by sentiment-based factors, and conducts additional tests with alternative machine learning methods. The results confirm that many factors, though potentially redundant for return prediction, play an important role in understanding pricing mechanisms and guiding investment decisions.