High-performance stock index trading via neural networks and trees

Abstract

Automated asset trading typically involves a price prediction model – of as high an accuracy as possible – together with a trading strategy, sometimes as simple as buying or selling when the price is predicted to rise or fall, respectively. Despite the fact that the model’s effectiveness in generating profits may depend on the particular trading strategy it is used with, these two components are often designed separately, in part because of the difficulty involved in jointly optimizing them. Motivated by this interplay between model performance and trading strategy, this work presents a novel automated trading architecture in which the prediction model is tuned to enhance profitability instead of accuracy, while the trading strategy attempts to be more sophisticated in its use of the model’s price predictions. In particular, instead of acting simply on whether the price is predicted to rise or fall we show that there is value in taking advantage of the model-specific distribution of predicted returns, and the fact that a prediction’s position within that distribution carries useful information about the expected profitability of a trade. Our proposed approach was tested with tree-based models as well as one deep long short-term memory (LSTM) neural networks, all of which were kept structurally simple and generated predictions based on price observations over a modest number of trading days. Tested over the period 2010–2019 on the S&P 500, Dow Jones Industrial Average (DJIA), NASDAQ and Russell 2000 stock indices, and our best overall model achieved cumulative returns of 350%, 403%, 497% and 333%, respectively, outperforming the benchmark buy-and-hold strategy as well as other recent efforts.