Factors Influencing the Predictive Performance of the LSTM Model on Stock Prices and Its Application in Forecasting U.S. Technology Sector Stock Prices — Sample Time Span, Time Window Length, Feature Selection, and Prediction Days

In recent years, the increasing global economic uncertainties have made financial markets more complex, posing challenges to stock price forecasting using traditional time series models. Therefore, optimizing the LSTM (Long Short-Term Memory) model and enhancing its prediction accuracy is of great significance for maintaining financial market stability. This paper primarily investigates the effects of four factors—sample time span, time window length, feature selection, and prediction days—on the predictive performance of the LSTM model. The study selects five representative stocks from the U.S. technology sector and employs the controlled variable method to separately alter these four factors. The variations in the model's mean squared error (MSE) were analyzed, and the underlying causes were hypothesized. The findings are as follows: First, an excessively small or large sample time span leads to a decline in the model's predictive ability; second, an overly long time window length results in reduced predictive performance; third, the impact of feature selection on the model's predictive ability varies significantly across different stocks; fourth, increasing the number of prediction days leads to a rise in the model's mean squared error; fifth, among these four factors, the sample time span has the most significant impact on the model's predictive ability.