Market Forecast using XGboost and Hyperparameters Optimized by TPE

Abstract

Online trading allows for thousands of transactions to occur within a fraction of a second, resulting in nearly unlimited opportunities to potentially find and take advantage of price differences in real time. However, in a fully efficient market, profit-oriented trading is a very important but difficult problem to solve. In this paper, in order to simplify the street trading problem, we propose to use the xgboost-based stock trading action selection prediction model and a special feature engineering process, as well as a hyperparameter optimization method. Our method can efficiently analyze attributes of different dimensions to make predictions better. We evaluated our XGboost trading behavior on the Jane Street dataset provided by the kaggle competition. Through the experiment result, our model shows surprising capability by contrast with other machine learning methods. Our profit indicators are 123 and 989 higher than those without hyperparameter optimization and neural network methods, respectively. In addition, we also studied the importance of features and hyperparameters.