A research on outlier in time series

SUMMARY The concept of outliers investigated in this study have been regarded as the observations which are extremely different from the others in a sample or one or two observations which effect the estimation and forecasting process in a great extent. The outliers may result from sevaral reasons such as; lack of necessary transformations, the errors made on observations and not being able to assume the correct model. In other words, some observations may be very "different" for different reasons other than the naturel randomness. The outliers in economic investigations may happen after expected or unexpected consequences. The data used in this study were obtained from the simulation experiments performed on several time series model by Chang, Tiao and Chen (1988) The original simulation results were reorganized and remodeled under a suitable experimental design. Two different time series models (AR, MA) with two outliers types (AO, 10) in two different size (3a, 5a) were arranged in three different time series size (50, 100, 150) and different sensitivity coefficient (3.00, 3.50, 4.00). The type of factorial design was (23 32) that is 3 factors each at 2 levels and two factors at 3 levels each. The result of the analysis of variance performed for this design indicated that the main effect except that of sensitivity coeffecient were statistically significant. On the other hand the iterative outlier search procedure was found to be most effective search process. A computer program was developed for this effective search procedure.