Forecasting Financial Stocks Using Data Mining

Abstract

This study presents a Business Intelligence (BI) approach to forecast daily changes in seven financial stocks’ prices from September 1, 1998 to April 30, 2008 with 267 independent variables. The purpose of our paper is to compare the performance of Ordinary Least Squares model and Neural Network model to see which model better predicts the changes in the stock prices and to identify critical predictors to forecast stock prices to increase forecasting accuracy for the professionals in the market. We used SPSS to perform stepwise regression to create a unique regression model for each company. Then, we ran the neural network with Alyuda NueroIntelligence to create a NN model by performing data analysis, data preprocessing, network design with hyperbolic tangent method, training with batch back propagation, testing, and query. We did data manipulation by using the first derivative and adding 0.1 to the absolute value of the minimum value in each variable to avoid minus sign from the rounding. Finally, we tested the model with the paired t-test in 152 randomly selected data points. Our result showed that the neural network model (batch back propagation algorithm) outperformed OLS model. The %error for NN and OLS mean ranges from 2.13%-3.27% and 4%-32% and standard deviation ranges from 1.78%-3.39% and 2.46%-8%. The OLS model is easy to use, validate, and works fast with lower forecasting accuracy because it is a linear model. NN has a better forecasting accuracy with no explanation of the relationship between interacting variables with dynamic results due to the learning setup. Some critical success factors to train NN are the network architecture, network design algorithm, training algorithm, and stop training conditions. Data normalization can make a huge difference to the result. We recommend more forecasting method and independent variables (e.g. expert opinion) to be included for future studies.