A Deep-Learning Approach in Asset-Pricing Anomalies

Project: Research

View graph of relations


Both academic and industry researchers keep finding tradable anomalies that produce excess returns. The increasing dimension of risk factors raises concerns among academics about data mining, data snooping, and the issue of multiple measurements. The financial statistics literature mainly focuses on factor testing for the risk premium. In this project, we take a predictive perspective and construct a multi-layer neural network to perform a sequential dimension process starting from the space of firm characteristics, and let the neural network to build its best tradable factors. We want to show the characteristics-based factor model, which begins since Fama and French (1993), is a shallow learner to a deep learning architecture. The asset pricing test tests the hypothesis that pricing errors, model implied expected returns minus realized returns, are jointly statistically and economically insignificant.However, after decades of factor fishing, there still exist significant pricing errors to challenge the asset pricing test in the sense of Gibbons, Ross, and Shanken (1989). We want to propose a data-driven optimization problem that minimizes the mispricings of the stock cross-section relative to a synthetic asset pricing model. To build a multi-layer neural network, we expect to explore a latent factor structure to perform a sequential dimension process from the space of firm characteristics to the best predictors, which is a regularization path to visualize a neural network factor model. We plan to provide an interpretable and scalable computational algorithm (TensorFlow) of asset pricing with two goals. One is to eliminate the model mispricings, namely, alphas, by detecting and exploiting those economically invisible nonlinearities and interactions in the data. The other is to illustrate a sequential dimension reduction process within a neural network, to which statistical factor models are shallow learners. 


Project number9048131
Grant typeECS
Effective start/end date1/01/19 → …