A Dynamic Feature Selection Technique for the Stock Price Forecasting

Abstract

Stock market prices are inherently volatile, and accurate forecasting is challenging. An accurate prediction of stock prices helps traders and investors to decide timely buy or sell, so an optimal investment strategy can be built, decreasing investment risks. Traditionally, linear and non-linear methods have been applied to stock market prediction. Many studies on stock market prediction have recently employed machine learning and deep learning models with the proliferation of big data and rapid development in artificial intelligence.

On the other hand, previous prediction studies mostly overlooked key indicators and feature engineering in the models. The feature selection can help to develop better prediction models. The stock price prediction requires a dynamic feature selection due to its time-dependent characteristics. There is no optimal set of technical indicators for stocks that perform well in all market scenarios. We propose a stock price prediction model focusing on dynamic feature selection in this study. The model uses technical, operational, and economic indicators besides price and volume data. The feature selection process has two stages. In the first stage, the importance of features for stocks is found by an ensemble learning algorithm. The final importance score is calculated by multiplying feature importance values with the next day’s model return which is the performance of the prediction method. In the second stage, a regression analysis is made daily for each feature using feature importance scores to track their performance in terms of average importance and slope (importance movement) dynamically. The proposed model enables better interpretability of features on stock price behavior and makes better stock price predictions.