While various linear and nonlinear forecasting models exist, multivariate methods like VAR, Exponential smoothing, and Box-Jenkins’ ARIMA methodology constitute the widely used methods in time series. This paper employs series of Turkish private consumption, exports and GDP data ranging between 1998: Q1 and 2017: Q4 to analyze the forecast performance of the three models using measures of accuracy such as RMSE, MAE, MAPE, Theil’s U1 and U2 . Seasonal decomposition and ADF unit root tests were performed to obtain new deseasonalized series and stationarity, respectively. Results offer preference for the use of ARIMA in forecasting, having performed better than VAR and exponential smoothing in all scenarios. Additionally, VAR model provided better forecast accuracy than exponential smoothing on all measures of accuracy except on Thiel’s U2 whose VAR values were not computed. Cautionary use of ARIMA for forecasting is recommended.
Keywords: Forecast Evaluation, ARIMA, Exponential Smoothing, VAR