Evaluating Performance of Inflation Forecasting Models of Pakistan

This study compares the forecasting performance of various models of inflation for a developing country estimated over the period of last two decades. Performance is measured at different forecast horizons (up to 24 months ahead) and for different time periods when inflation is low, high and moderate (in the context of Pakistan economy). Performance is considered relative to the best amongst the three usually used forecast evaluation benchmarks – random walk, ARIMA and AR(1) models. We find forecasts from ARDL modeling and certain combinations of point forecasts better than the best benchmark model, the random walk model, as well as structural VAR and Bayesian VAR models for forecasting inflation for Pakistan. For low inflation regime, upper trimmed average of the point forecasts out performs any model based forecasting for short period of time. For longer period, use of an ARDL model is the best choice. For moderate inflation regime different ways to average various models’ point forecasts turn out to be the best for all inflation forecasting horizons. The most important case of high inflation regime was best forecasted by ARDL approach for all the periods up to 24 months ahead. In overall, we can say that forecasting performance of different approaches is state dependent for the case of developing countries, like Pakistan, where inflation is occasionally high and volatile.