Details of Research Outputs

Outliers are commonplace in many real-life experiments. The presence of even a few anomalous data can lead to model misspecification, biased parameter estimation, and poor forecasts. Outliers in a time series are usually generated by dynamic intervention models at unknown points of time. Therefore, detecting outliers is the cornerstone before implementing any statistical analysis. In this paper, a multivariate outlier detection algorithm is given to detect outliers in time series models. A univariate time series is transformed to bivariate data based on the estimate of robust lag. The proposed algorithm is designed by using robust measures of location and dispersion matrix. Feed forward neural network is used for designing time series models. Number of hidden units in the network is determined based on the standard error of the forecasting error. A comparison study between the proposed algorithm and the widely used algorithms is given based on three real-data sets. The results demonstrated that the proposed algorithm outperformed the existing algorithms due to its non-requirement of a priori knowledge of the time series and its control of both masking and swamping effects. We also discussed an efficient method to deal with unexpected jumps or drops on share prices due to stock split and commodity prices near contract expiry dates.