Multilayer Perceptron Neural Network Models in Asset Pricing: An Empirical Study on Large-Cap US Stocks

In this study, MLP models with dynamic structure are applied to factor models for asset pricing tasks. Concretely, the MLP pyramid model structure was employed on firm characteristic-sorted portfolio factors for modelling the large-cap US stocks. It was further developed as a practical factor investing strategy based on the predictions. The main findings were evaluated from 2 angles: model predictive power and backtesting performance, which were compared for the periods with and without COVID-19. The empirical results indicated that, given the constraints of the data size, the MLP models no longer perform 'deeper, better' in terms of predictive power, whereas the proposed MLP models with 2 and 3 hidden layers have greater flexibility in modelling the factors in this case. This study also verified the idea from previous work that MLP models for factor investing are more meaningful for downside risk control than for pursuing absolute annual returns.