Asset Pricing in Neural Network Models

The thesis investigates the application of advanced deep learning methods to predict stock returns within the framework of empirical asset pricing. This study explores how traditional and advanced neural network models can improve the predictive accuracy of stock excess returns and to what extent these models can price them. Concretely, it applies neural network frameworks, such as Multilayer Perceptron (MLP), recurrent neural network (RNN) and its variations, RNN attention models and Transformer models, as substitutes for the traditional linear framework of the asset pricing factor model. These neural network models are tested on monthly data for 420 large-cap US stocks from January 1957 to December 2021 (Chapter 2) and December 2022 (Chapter 3 and 4). This extensive sample enables a robust assessment across diverse market regimes, capturing idiosyncratic volatility clusters and structural shifts during the COVID-19 pandemic and its subsequent recovery. The 182 firm characteristics-sorted factors explored by previous researchers are selected from the literature.

Chapter 2 develops dynamic pyramidal MLP structures for factor models that learn nonlinear mappings from factors to stock excess returns. The MLP models with 2 hidden layers outperform alternative MLP and traditional linear models on predictive power. In backtesting, with the dynamic transaction cost scenario, the deep MLP models slightly outperform the buy-and-hold benchmark in both portfolios and periods. The results also prove the capability of MLP models in moderating the extreme downside risks. Chapter 3 proposed two tailored innovative pre-trained RNN attention structures for stock return prediction under the empirical asset pricing framework, which are the pre-trained RNN global self-attention model and the sliding window sparse attention model. They are benchmarked with the pre-trained RNN models and their variants (LSTM and GRU) and pre-trained RNN attention models (additive, Luong's three attentions). The out-of-sample R-Square shows the RNN model has significant advantages in modelling the stock return movements during the extreme market fluctuations, which is followed by the proposed models. Nonetheless, the backtesting results show that the proposed models outperform all alternative benchmarks on profitability in both equal-weighted and value-weighted portfolios. Chapter 4 proposes an innovative Transformer model, Single-directional representative from Transformer (SERT). It applies pre-trained Transformer models in the context of stock pricing. They are compared with standard Transformer models and encoder-only Transformer models across three periods. The best proposed SERT model achieves the highest out-of-sample R-Square during the extreme market shocks, followed by pre-trained Transformer models. Their backtesting performance proves the excellent capability for hedging extreme downside risks. It also proves that Transformer models have high capabilities to capture patterns of highly volatile temporal sparsity data in the asset pricing context. The dynamic transaction cost robustness examination shows that the profitability of these models is highly eroded by the high turnover rate.