Feature Selection Based on Stock Prediction Model

Abstract

Stocks as an important part of financial investment are becoming more and more popular, and they have higher rates of both returns and risks. Making a prediction for the stock can reduce its risk and help people gain returns. So far, the traditional machine learning model is still unable to achieve ideal accuracy. The paper is devoted to analyzing the input features to improve the performance of stock forecasting models. Aiming at the problem that the traditional stock prediction algorithms produce different accuracy for the models constructed by different input features, the paper, through a method of establishing a long-term memory (LSTM) model, predicts the stock. The Shixia Technology Stock's history data includes 5 features as the dataset and chooses the different feature as the input. In the experiment in this paper, different results are produced by subtracting one different input feature at a time. Finally, the model's predictions were compared with each other by the R_square and RMSE, and the analysis revealed which feature could have a greater impact on the stock prediction. The paper finds that the different input features have different influences on the model fitting effect and the prediction accuracy on stock forecasting based on the same neural network model.