Stacked Model with Autoencoder for Financial Time Series Prediction

Abstract

In this paper, we propose a stacked model with autoencoder for financial time series prediction. A stacked autoencoder model is used for feature extraction of high-dimensional stock factors. The factors after dimensionality reduction serve as input to the stacked model to predict the next-day returns of the stocks. In this paper, the stacked autoencoder not only has the effect of reducing the dimension, but also eliminates the redundant information in the data to a certain extent, which can effectively improve the predictive capacity of the model. The constituent stocks of CSI300 are used as backtest samples, and the experiment shows that the stacked model with autoencoder can obtain more than 50% of excess return in 2019.