Enhanced Hybrid Prediction Models for Time Series Prediction

Statistical techniques have disadvantages in handling the non-linear pattern. Soft Computing (SC) techniques such as artificial neural networks are considered to be better for prediction of data with non-linear patterns. In the real-life, timeseries data comprise complex pattern, and hence it may be difficult to obtain high prediction accuracy rates using the statistical or SC techniques individually. We propose two enhanced hybrid models for time series prediction. The first model is an enhanced hybrid model combining statistical and neural network techniques. Using this model, one can select the best statistical technique as well as the best configuration for the neural network for time series prediction. The second model is an enhanced adaptive neuro-fuzzy inference system which combines fuzzy inference system and neural network. The proposed enhanced Adaptive Neuro-Fuzzy Inference Systems (ANFIS) model can determine the optimum input lags for obtaining the best accuracy results. The prediction accuracies of the two proposed hybrid models are compared with those obtained with other models based on three time series data sets. The results indicate that the proposed hybrid models yield better accuracy results compared to Autoregressive Integrated Moving Average (ARIMA), exponential smoothing, moving average, weighted moving average and Neural Network models.