Stock market forecasting using a multi-task approach integrating long short-term memory and the random forest framework

Abstract

Numerous studies have adopted deep learning (DL) in financial market forecasting models owing to its superior performance. The DL models require as many relevant input variables as possible to improve performance because they learn high-level features from these inputs. However, as the number of input variables increases, the number of parameters also increases, leaving the model prone to overfitting. We propose a novel stock-market prediction framework (LSTM–Forest) integrating long short-term memory and random forest (RF) to address this issue. We also develop a multi-task model that predicts stock market returns and classifies return directions to improve predictability and profitability. Our model is interpretable because it can identify key variables using the variable importance analysis from RF. We used three global stock indices and 43 financial technical indicators to verify the proposed methods experimentally. The root mean squared errors of LSTM–Forest with multi-task (LFM) in predicting returns from S&P500, SSE, and KOSPI200 were 25.53%, 22.75%, and 16.29% lower, respectively, than those of the baseline RF model. The model’s balanced accuracy in forecasting the daily return direction increased by 7.37, 1.68, and 3.79 points, respectively. Furthermore, our multi-task model outperforms our single-task model and previous DL approaches. In the trading test, LFM produced the highest profits—even compared with the long-only strategy when transaction costs were considered. The proposed framework is readily extensible to other tasks and fields that contend with high-dimensionality problems.