Abstract: As structural breaks and outliers exist in complex time series data, they may easily affect the training process of a forecasting model, and the forecasting model can generate extreme forecasting results. In this case, a trimmed average is proposed based neural network ensemble forecasting approach. This approach first generates multiple training sets, then the training sets are used to training multiple neural network model. Finally, the forecasting results are combined using trimmed average strategy. Compared with simple average, trimmed average is less sensitive to outliers, and makes the ensemble model has better forecasting accuracy. In the empirical study, two neural network ensemble models are built, namely trimmed average based bootstrap neural network ensemble (TA-BNNE) and trimmed average based Monte Carlo neural network ensemble (TA-MCNNE). These two models are used to forecast the NN3 competition datasets. The results show that they both have better forecasting accuracy than simple average based ensemble models. Moreover, the performance of our proposed model is compared with those that ranks the first 10 place in NN3 competition in the competition context, and their performance is found both surpass that of the 6th ranking model for all datasets and 1st ranking model for complex datasets, indicating that our proposed approach is very effective. Our approach extends the trimmed average strategy to time series forecasting and shows promising forecasting accuracy especially for time series with structural breaks and outliers. Moreover, this approach can be easily extended to different time series forecasting applications, since neural network ensemble as a popular forecasting tool has been shown to be very effective for time series forecasting in an extensive body of literatures.

Key words: time series forecasting, ensemble model, neural network, trimmed average, resampling