Rules extraction in short memory time series using genetic algorithms

Data mining is performed using genetic algorithm on artificially generated time series data with short memory. The extraction of rules from a training set and the subsequent testing of these rules provide a basis for the pred ictions on the test set. The artificial time series are generated using the inverse whitening transformation, and the correlation function has an expo nential form with given time constant indicative of short memory. A vector quantization technique is employed to classify the daily rate of return of this artificial time series into four categories. A simple genetic algorith m based on a fixed format of rules is introduced to do the forecasting. Com paring to the benchmark tests with random walk and random guess, genetic al gorithms yield substantially better prediction rates, between 50% to 60%. T his is an improvement compared with the 47% for random walk prediction and 25% for random guessing method.