An Information Theoretic Approach to Market Index Prediction

This paper compares information theoretic approaches to building Bayesian belief networks to perform market index prediction. We show that the automatic model building can be done efficiently using the 2 ! criteria rather than mutual information alone. We suggest that when the number of variables are small, and instantiations are small, this criteria is a straightforward way of determining conditional independence. Both approaches use identical data to predict the stock market returns as a function of macroeconomic variables and the results are comparable (62% vs. 63% accuracy). We discuss the relative advantages of belief networks.