Forecasting Performance of Arma and Arfima Models in Short Time Series: An Analysis of Kenya Political Opinion Poll Data
Otieno M. O., Mwangi J. W., Islam A.S.

Abstract
The popular approach to modeling time series data is to apply the Box-Jenkins approach of ARMA or ARMA depending on whether the series is stationary or non-stationary. This approach is based on the assumption that the sample is large. If such series display long memory property, then the forecast values based on ARMA model may not be reliable. In case of short time series data, one cannot rely on estimation techniques based on the asymptotic theory. This calls for use of appropriate estimation techniques in order to come up with models that can capture the short time series properties and thus be adequately used for prediction and forecasting without loosing the principle of parsimony. This study focuses on fitting appropriate ARMA and ARFIMA models for short time series and measuring the forecast performance of the fitted models. The percentage political popularity ratings series for three presidential candidates in Kenya’s General Elections for the year 2007 are used. A model-selection strategy based on the corrected Akaike Information Criterion (AICc) is adopted to determine the correct model specification. Exact maximum likelihood estimation method is used to estimate the model parameters. RMSE is used to evaluate the forecast performance of the model. ARFIMA models are found to represent and forecast the short time series polls data better than the ARMA models.

Full Text: PDF