Seasonal Time Series Forecasting by Group Method of Data Handling

This paper uses Group Method of Data Handling (GMDH) as a self organizing data mining technique based on automatic generation of optimum multilayer polynomial network structures that can extract knowledge about an object directly from data samples. It finds the best model by sorting-out of possible variants and minimizes the influence of the design parameters on the results of modelling. This machine learning approach optimises the model structure and laws that govern the input output relationships of a system. Unlike neural networks, the number of nodes and the model structure are identified automatically. GMDH has been employed to forecast the standard time series of airline passenger data and results have been compared with the traditional neural networks and Box & Jenkins Auto Regressive Integrated Moving Average (ARIMA) model to testify the superiority of the proposed technique. A study of the effect of different ratios of training to testing data on the proposed model has also been carried out.