Prediction-based Stock Portfolio Optimization Using Bidirectional Long Short-Term Memory (BiLSTM) and LSTM

Abstract

Purpose: Investment is the allocation of funds with the aim of obtaining profits in the future. An example of the investment instruments with high returns and high risks are stocks. The risks associated with the investment can be reduced by forming a portfolio of quality stocks optimized through mean-variance (MV). This is necessary because successful selection of high-quality stocks depends on the future performance which can be determined through accurate price prediction.

Methods: Stock price can be predicted through the adoption of different forms of deep learning methods. Therefore, BiLSTM and LSTM models were applied in this research using the stocks listed on the LQ45 index as case study.

Result: The utilization of LSTM and BiLSTM models for stock price prediction produced favorable outcomes. It was observed that BiLSTM outperformed LSTM by achieving an average MAPE value of 2.1765, MAE of 104.05, and RMSE of 139.04. The model was subsequently applied to predict a set of stocks with the most promising returns which were later incorporated into the portfolio and further optimized using the Mean-Variance (MV). The results from the optimization and evaluation of the portfolio showed that the BiLSTM+MV strategy proposed had the highest Sharpe Ratio value at k=4 compared to the other models. The stocks found in the optimal portfolio were BRPT with a weight of 19.7%, ACES had 16.9%, MAPI 11.8%, and BMRI at 51.6%.

Novelty: This research conducted a novel comparison of LSTM and BiLSTM models for the prediction of stock prices of companies listed in the LQ45 index which were further used to construct a portfolio. Past research showed that the development of portfolios based on predictions was not popular.