Machine Learning Models' Combination for Higher Accuracy of S&P 500 Index Prediction

This paper addresses problems of predicting the intraday price flow in S&P 500 index. The study compares the accuracy and profit rate of each single and different combinations of models of classifiers: Fully-connected Neural Network, Support Vector Machine, and linear regression with basic buy-and-hold strategy. The inputs adopt technical indicators from the raw data. The learning is carried out on 10 years of historical data from 2002 to 2012, and the back test is conducted on a 4-year-period from 2013 to 2017. The result of prediction in back test has the accuracy over 63% of the time and earns 37.5% annual profit. The experimental results show that the hybrid of classifiers provides higher accuracy on price flow trends, but the frequency of trading decreases, while for particular time period single classifier may yield higher profit. In conclusion, this experiment aims to improve the accuracy and profit of the machine learning model that predicts the intraday price of S&P 500 index through different means of combination of existing models.