Combined group method of data handling models using artificial bee colony algorithm in time series forecasting

摘要

Time series forecasting is an important area of forecasting which has gained many attentions from various research areas. In line with its popularity, various models have been introduced for the purpose of producing accurate time series forecasts. Nevertheless, it is difficult to find an ideal model as there is no model that can perform best for all types of data. Recently, combination of forecasts has gained immense popularity in the time series forecasting area. Among the well-known combination technique is the weighted-based approach, where appropriate weights are given based on the performance of each individual model. In this paper, a robust methodology for accurate time series forecasting based on Artificial Bee Colony (ABC) algorithm and four Group Method of Data Handling (GMDH) models, namely GMDH-Polynomial, GMDH-Radial Basis Function, GMDH-Sigmoid and GMDH-Tangent are proposed. In this methodology, the first step was to develop forecasts using individual GMDH models. In the second step, weights for each individual GMDH were calculated heuristically using ABC algorithm. In the final step, the results were aggregated to form a new forecast. For the purpose of evaluating the performance of the proposed model, the model was applied to a real time series data which is the monthly tourist arrivals from Singapore to Malaysia from year 2000 to 2017. Based on the empirical results, the application of nonlinear transfer function such as Tangent function has the potential to improve the performance of GMDH model, while the other functions such as Sigmoid function produced a forecast which is worse than the conventional GMDH-Polynomial. However, the combination of GMDH models using ABC managed to outperform all the individual models. The empirical finding demonstrated the efficiency of Combined GMDH model in time series forecasting, as well as the applicability of implementing optimization algorithm to find appropriate weights for the individual models.