Return Prediction Based on Discriminating market-styles with Reinforcement Learning

For hedge funds, return prediction has always been a fundamental and important problem. Usually, a good return prediction model directly determines the performance of a quantitative investment strategy. However, the performance of the model will be influenced by the market-style. Even the models trained through the same data set, their performance is different in different market-styles. Traditional methods hope to train a universal linear or nonlinear model on the data set to cope with different market-styles. However, the linear model has limited fitting ability and is insufficient to deal with hundreds of features in the hedge fund features pool. The nonlinear model has a risk to be over-fitting. Simultaneously, changes in market-style will make certain features valid or invalid, and a traditional linear or nonlinear model is not sufficient to deal with this situation. This thesis proposes a method based on Reinforcement Learning that automatically discriminates market-styles and automatically selects the model that best ts the current market-style from sub-models pre-trained with different categories of features to predict the return of stocks. Compared with the traditional method that training return prediction model directly through the full data sets, the experiment shows that the proposed method has a better performance, which has a higher Sharpe ratio and annualized return.