Study of asset pricing: insights from factor models and dynamic methods

This thesis examines various aspects of the Fama-French five-factor model (FF-5) through different methodologies, including deep learning models, nonparametric methods, and traditional FF-5 model with liquidity factor. The 1st empirical chapter examines the Fama-French five-factor model’s ability to explain stock returns in Chinese market, focusing on the size effect and liquidity effect. It also analyses the relationship between idiosyncratic volatility and expected stock returns. The study finds that liquidity effects are weaker than size, value, and investment effects, with the six-factor model outperforming the five-factor model in certain portfolios. Additionally, it finds a negative correlation between idiosyncratic volatility and stock returns. The second study uses a nonparametric method to estimate the conditional factor model, offering an alternative to traditional linear models. By applying nonparametric approach, the study estimates the optimal bandwidth for conditional alphas and betas and compares high-frequency and low-frequency factors. The results indicate that the nonparametric method outperforms the OLS method in reducing pricing errors, making it a valuable tool for high-frequency empirical applications. The third research explores the application of the Fama-French five-factor model in financial forecasting, utilising deep learning techniques, specifically CNN, LSTM, CNN-LSTM, and TCN-LSTM models, to predict excess returns based on high-frequency data. Among the models, CNN-LSTM achieves the best performance, effectively balancing feature extraction and modelling.