A comparison of the forecasting ability of ARIMA models

Abstract

PurposeARIMA models have been extensively examined in the context of the real estate market. The purpose of this paper is to examine issues relating to their application in a forecasting context. Specifically, the paper seeks to examine whether in‐sample measures of best‐fit and also past forecasting accuracy bear any relation to future forecasting performance.Design/methodology/approachThe forecasting performance of alternative ARIMA specifications are compared over rolling estimation and forecasting windows. The forecasting accuracy of the alternative specifications is compared with specific attention placed on the accuracy of the respective specification that in‐sample provides the best fitting model.FindingsThe results highlight the limitations in using the conventional approach to identifying the best‐specified ARIMA model in sample, when the purpose of the analysis is to provide forecasts. The results show that while ARIMA models can be useful in anticipating broad market trends, there are substantial differences in the forecasts obtained using alternative specifications. The use of conventional measures of best‐fit provide little indication as to future forecasting ability, nor does the forecasting performance of a specification in previous periods.Originality/valueARIMA modelling has frequently been highlighted as a useful forecasting approach. This paper illustrates that care needs to be paid in their use in a forecasting context and full appreciation of the strengths and limitations of the ARIMA approach.