A Modified LSTM Neural Network Based on Boosting and Sufficient Dimensionality Reduction

Abstract

Stock price prediction is a common scenario in time series data forecasting, providing effective guidance for investment decisions. Long Short-Term Memory (LSTM) is a widely used model for stock price prediction, yet the selection of its hyperparameters remains an unresolved issue. In this paper, we address this challenge by employing model averaging instead of model selection. Specifically, we adaptively solve the hyperparameter selection problem by utilizing a distance covariance-weighted method, effectively balancing the bias and variance of the predictive model. Additionally, we propose an enhanced model that employs a boosting approach based on sufficiently reducing dimensionality through a multifactor model. This approach captures stock price sequence information beyond volatility. Practical data analysis demonstrates that the proposed method exhibits significant advantages over the original LSTM model in terms of mean square error or absolute error. Furthermore, the proposed framework can be applied to hyperparameter selection in other time series prediction models, such as autoregressive integrated moving averages (ARIMA), including the selection of autoregressive and partial autocorrelation orders.