Relational Feature Selection and Uncertainty Quantification for Financial Time Series Forecasting

Abstract

The contribution of this study is to improve the success of stock forecasting and to perform reliability analysis through uncertainty estimation. In the first stage of the study, a common change matrix was created by examining the directional movement of stocks in the same financial sector, and the most correlated stocks for a selected stock were determined using association rule analysis (ARM). A model was designed to predict the closing prices of the selected main stock from t+1 to t+5 days, and not only its own historical values but also the stocks selected through association analysis were used as input signals. Dual Attention Networks (DAN) were used in time series modeling. In the second stage of the study, uncertainty analysis was performed on the forecasting models using the Monte Carlo Dropout (MCD) method, and the points where uncertainty changes in forward forecasts were marked. The proposed methods were validated with a stock selected from the BIST database.