An Analysis of the Hybridization of a Divide & Conquer Strategy with Genetic Fuzzy Set for Financial Prediction

Abstract

Several works within the field of artificial intelligence & financial prediction show increased prediction performance by the hybridization of an existing prediction model with the division of the feature space into a number of regions prior to the development of separate prediction models for each region, herein referred to as a divide and conquer strategy. In this thesis comparisons are made between a genetic fuzzy set prediction model with and without the hybridization of the divide and conquer strategy using Kohonen Self-organizing Maps (SOM) as a classifier. The data used in this thesis consist of monthly cross sections of about 1400 companies per cross section with a set of 14 features of both fundamental and technical classifications. Comparisons are made of 288 different configurations differing in i) magnitude of data used in training of the model by the genetic algorithm (GA), ii) length of training period proceeding test period, iii) length of prediction horizon and iv) test periods. Analysis is done to assess the conditions of which the divide and conquer strategy increases performance relatively both in terms of 1) mean absolute percentage error (MAPE) and 2) mean return of top 10% of ranking companies (MRC) in 12 month test periods. Results suggests increase in performance with the divide and conquer strategy is A) correlated positively with longer prediction horizons for both MAPE and MRC, B) correlated positively with higher magnitudes of data for MAPE, but not correlated for MRC and C) not correlated with length of training period.