Inductive Logic Programming for Discovering Financial Regularities

The purpose of this work is discovering regularities in financial time series using Inductive Logic Programming (ILP) and related "Discovery" software system [Vityaev et al., 1992,1993] in data mining. Discovered regularities were used for forecasting the target variable, representing the relative difference in percent between today's closing price and the price five days ahead. We describe the method, types of regularities found and analyzed, statistical characteristics of these regularities on the training and test data and the percentage of true and false predictions on the test data. There are more than 130 discovered regularities on 10 year (1985-1994) data. The best of these regularities had shown about 75 % of correct forecasts on test data (1995-1996). The target variable was predicted using separately SP500 (close) and own history of the target variable. Active trading strategy based on discovered rules outperformed buy-and-hold strategy and strategies based on several ARIMA models in simulated trading for 1995-1996. An ARIMA model constructed using discovered rules had shown the best performance among tested ARIMA models. The performance of this model is similar to performance based on discovered rules.