Learning from Errors While Forecasting Inflation: A Case for Intercept Correction

Structural changes are quite common in macroeconomic time series. Moreover, any underlying macroeconomic relationship cannot be correctly specified unless we know the true model. Structural changes in time series and misspecification in empirical model are observed as shifts in the constant of the underlying relationship between the subject variables of interest. Forecasting from such a model assuming 'no structural break' and 'correct model' is tantamount to ignoring important aspects of underlying economy and mostly results in forecast failure(s). Intercept correction (IC) is a method for accommodating such ignored structural break(s) and omitted variable(s). We use a simple model (for July 1991 to March 2016) to forecast inflation for 25 countries and compare its performance with a) the same model with optimal IC, b) the same model with half IC, and c) a random walk model. Optimal IC approach, though computational intensive, outperforms in forecasting next period inflation compared to one from a) the same model without IC, b) the same model with half intercept correction, and c) random walk model without IC. For the particular class of inflation models under study, over the time period specified, 'quarter IC' works best among the fixed IC rules.