LSTMによる時系列予測と株式投資戦略への応用

抄録

Researchs for financial time series in stock or foreign exchange markets, have been one of traditional themes of financial market analysis. Statistical model approaches such as ARMA and GARCH were mainstream of conventional analysis. However, it is difficult to understand and predict financial time series structures, which are generally characterized by high noise level and low autocorrelation. Meanwhile, researchs to capture the structure by artificial intelligence has been increasing in recent years. In particular, Long Short-Term Memory (LSTM), which can capture time series structure, is already widely used in the field of natural language processing and speech recognition. Therefore, in this study, we investigated the model performance in each TOPIX core30 constituent stock by using logistics regression (LOG), random forest (RAF), gradient boosting (GBT), support vector machine (SVM), and LSTM. The performance was evaluated by metrics such as prediction accuracy, F1 measure, AUC, and return. As a result, LSTM showed the best performance in the models. Moreover, we discussed the effectiveness of the stock market neutral strategy by applying the above prediction models. 10-quantile portfolios using the predicted probability outputted by the model, remarks higher accuracy and returns than individual stock trading in all models. Furthermore, LSTM outperformed the others and it is consistent with the result of S&P500 constituent stocks analysis.