Multiobjective optimization based adaptive models with fuzzy decision making for stock market forecasting

Stock market forecasting is an important and challenging task. Conventional single objective optimization based adaptive prediction models reported in the literature do not satisfy many cost functions simultaneously. Very few reported materials are available on the development of multiobjective optimization based stock market prediction models. In this paper multiobjective particle swarm optimization (MOPSO) and nondominated sorting genetic algorithm version-II (NSGA-II) have been introduced to effectively train the adaptive stock market prediction models which simultaneously optimize four performance measures. The model developed is an adaptive one with nonlinearity introduced at the input end by Legendre polynomial expansion scheme. The stepwise algorithms are provided to develop the model and simulation study is carried to evaluate the performance. To arrive at the best possible solutions from these models, fuzzy logic based decision making strategy is suggested. Close examination of simulation results reveals that in terms of directional accuracy (DA) and computation time MOPSO based model is better where as in terms of average relative variance (ARV) and I-metric the NSGA-II model is superior. However, with regard to mean average percentage of error (MAPE) and Theli's U, MOPSO is better above one month ahead prediction. But for below one month ahead prediction, the NSGA-II model is preferable. To facilitate comparison two single objective optimization based models (PSO and GA based) are also developed and the performance has been obtained through simulation study. Comparison of the results demonstrate that in terms of MAPE and DA the performance of multiobjective is better where as the single objective optimization model exhibit superior performance in terms of Theli's U and the ARV.