Evolution of high-frequency systematic trading: a performance-driven gradient boosting model

This paper proposes a performance-driven gradient boosting model (pdGBM) which predicts short-horizon price movements by combining nonlinear response functions of selected predictors. This model performs gradient descent in a constrained functional space by directly minimizing loss functions customized with different trading performance measurements. To demonstrate its practical applications, a simple trading system was designed with trading signals constructed from pdGBM predictions and fixed holding period in each trade. We tested this trading system on the high-frequency data of SPDR S&P 500 index ETF (SPY). In the out-of-sample period, it generated an average of 0.045% return per trade and an annualized Sharpe ratio close to 20 after transaction costs. Various empirical results also showed the model robustness to different parameters. These superior performances confirm the predictability of short-horizon price movements in the US equity market. We also compared the performance of this trading system with similar trading systems based on other predictive models like the gradient boosting model with L2 loss function and the penalized linear model. Results showed that pdGBM substantially outperformed all other models by higher returns in each month of the testing period. Additionally, pdGBM has many advantages including its capability of automatic predictor selection and nonlinear pattern recognition, as well as its simply structured and interpretable output function.