Portfolio Construction with Stock Prices Predicted by LSTM using Enhanced Features

Abstract

Machine learning methods, which have shown great success in financial applications, have also been effective in portfolio construction recently. Although raw time series data are generally used in these studies, recent studies have shown that features extracted from time series data increase the capacity of the models. In this study, we propose a portfolio creation pipeline that utilizes features extracted from images generated from stock data. First, we generate a candlestick chart from the raw stock data and extract semantic information from them with the help of an autoencoder. Next, we train a long short-term memory (LSTM) model that predicts stock prices with enhanced features which are the combination of extracted features with raw stock data. Portfolios that are estimated to bring the highest return are constructed by using the estimated price information. We evaluate the performance of the portfolio construction method with stock data in Standard & Poor 500 index in terms of daily mean return, Sharpe ratio, and compound return. We compare with three baselines portfolio construction methods: 1/N rule, portfolio construction by Sharpe ratio and portfolio construction by LSTM that takes raw time series as input. Our model outperforms 1/N, Sharpe ratio, and LSTM that uses raw time series.