The Portfolio Model Based on Temporal Convolution Networks and the Empirical Research on Chinese Stock Market

Abstract

Aiming at determining the weights of selected investment targets to obtain higher and more stable investment returns, a Temporal Convolution Networks (TCN) based portfolio model, namely, TCNportfolio model is proposed. TCN-portfolio model combines TCNbased time series processing and MLP-based cross-sectional data processing, and finally outputs the investment target weights which changes every ten trading days. We optimize the TCN-portfolio model using a Multi-Objective Genetic Algorithm (MOGA) which optimizes the rate of return and variance at the same time. The component stocks of Shanghai Securities Composite 50 index (SSEC 50) are selected as the investment targets. Experimental results on the test sets reveal that TCN-portfolio model performs well. Its average daily return rate is obviously greater than those of SSEC and SSEC 50, and the cumulative return rate of TCN-portfolio model is always greater than those of SSEC and SSEC 50 on the test data set.