Optimizing Long Short -Term Memory (LSTM) Model Hyperparameters for Enhanced Stock Price Forecasting and Portfolio Allocation

Due to the increasing application of Long Short-Term Memory (LSTM) models in stock price forecasting and portfolio allocation, it is crucial to tune the models for better accuracy. However, there is a limited study on how the hyperparameters of LSTM models affect the model performances. Therefore, this paper investigated the relationships between hyperparameters, particularly the number of neurons and LSTM layers, on the Mean Squared Error (MSE) of LSTM models. To shed light on practical significance, this research was conducted in the setting of portfolio optimization with a combination of LSTM stock price forecasting and Monte-Carlo Portfolio Simulation. More specifically, the LSTM model was first trained to forecast the weekly prices of five selected stocks, during which the MSE resulted from different number of neurons in the first LSTM layer as well as the total number of layers were compared. Following that, the combination of hyperparameters reaching the smallest MSE was selected for each stock, and the corresponding forecasted returns (calculated from the forecasted prices) were treated as input of the Monte-Carlo Simulation. Finally, the Monte-Carlo Simulation was employed for generating the desired portfolios. As the results demonstrated, the increase in number of layers in general leads to rising MSE, yet the increase in number of neurons in the first LSTM layer has either blurred effect or improving effects on model performances, depending on the original volatility of historical values.