How well do linear and nonlinear time series models’ forecasts compete with international economic organizations?

Abstract

This paper evaluates the forecasting performance of linear and non-linear time series models of some macroeconomic variables viz a viz the forecasts outlook of these variables generated by professionals in international economic organizations i.e. the International Monetary Fund (IMF) and the Organization of Economic Cooperation and Development (OECD). Many time series and econometrics models are used to forecast financial and macroeconomic variables. The accuracy of such forecasts depends crucially on careful handling of nonlinearity present in the time series. The debate of forecasting ability of linear vs nonlinear models is far from settled. These models use the past patterns of the economic time series to infer the parameters of the underlying stochastic process and use them to make forecasts. In doing so these models use only the information contained in the past data. However the economists working in professional international economic organizations not only look at the past trends but use the condition of local and global economy prevailing at the time and expected future path of economies as well as their professional expertise and judgment to arrive at forecasts of macroeconomic variables. However the specific underlying models and methodology used by the economists generating these forecast is usually not communicated to the public. In comparison to the forecasts of these organizations the time series models are well developed and accessible to researchers working anywhere around the globe. Thus it is an interesting task to compare the foresting ability of linear and nonlinear time series models. This paper aims at comparing the forecasts from these models to assess how well they compete with forecasts generated from the professional economists employed by international economic organizations. The nonlinear models employed in this study are quite well known namely the Self Exciting Threshold Autoregressive (SETAR) model and the Markov Switching Autoregressive (MSAR) model. The linear models employed are the AR and ARMA models. The paper have used annual data of three macroeconomic time series variables GDP growth, consumer price inflation and exchange rate of G7 countries i.e. Canada, France, Germany, Italy, Japan, United Kingdom (UK) and United States of America (USA) as well as an emerging south Asian economy namely Pakistan. Three forecast accuracy criteria i.e. Root Mean Square Error (RMSE), Mean Absolute Error (MAE) and Mean Absolute Percentage Error (MAPE) are employed and the statistical significance of difference in forecasts is assessed using the Diebold-Mariono test. The results show that the forecasting ability of nonlinear Regime Switching models SETAR and MSAR is superior to the linear models. Further, although the point forecasts of linear and nonlinear models are not superior to that of economic organizations but in more than 60 percent of the cases considered the forecasting accuracy of two sets of forecast is not statistically significantly different.