Performance of ARIMA models in time series

This study is mainly concerned with an evaluation of the forecasting performance of a set of the most often applied ARIMA models. These models were fitted to a sample of two hundred seasonal time series chosen from eleven sectors of the Canadian economy. The performance of the models was judged according to eight variable criteria, namely: average forecast error for the last three years, the chi-square statistic for the randomness of the residuals, the presence of small parameters, overdifferencing, underdifferencing, correlation between the parameters, stationarity and invertibility. Overall and conditional rankings of the models are obtained and graphs are presented.