A D-GMDH model for time series forecasting

Traditional GMDH (group method of data handling) method has been applied in series forecasting successfully many times. In this paper, we bring concept of diversity into GMDH to improve the noise-immunity ability. Five diversity metrics are used as external criteria to construct a new kind of GMDH forecasting models called D-GMDH. To assess the effectiveness of D-GMDH, we compare them with traditional GMDH method, autoregressive integrated moving average (ARIMA) and artificial neural network (ANN), and find out that the two models – D-GMDH (chi) and D-GMDH (cor) – are better than the others among the five D-GMDH models. The two better models are then used to forecast financial time series with noise. Results show that the two new proposed models can provide high forecasting accuracy in noisy environment.