Stock return forecasting based on the proxy variables of category factors

Abstract Stock return prediction has been in the spotlight because it involves numerous factors. Improving the accuracy of stock return prediction and quantifying the impact of individual factors on forecasting remain challenging tasks. Motivated by these challenges, we propose a novel forecasting method that entails proxy variables of category factors and the random forest technique. This new method aims to quantify the information and importance of category factors, thereby enhancing the predictability of stock returns. Specifically, we categorize a large set of return predictors into several category factors. We then utilize the importance of the original variables to construct proxy variables for these category factors. Subsequently, we use the proxy variables to build a random forest model for predicting stock returns. Our empirical analysis results demonstrate that the proposed method effectively quantifies the importance of both the original factors and category factors. Furthermore, we find that the fundamental information factor consistently ranks as the most crucial category factor for stock return forecasting. Additionally, the proposed method exhibits a more robust and prominent prediction performance than competing models such as single-category-factor-based random forest models, dimension-reduction, and forecast-combination methods. Most importantly, the proposed method produces forecast results that can assist investors with understanding stock market dynamics and facilitate higher investment returns.