A Statistical Forecasting Method for Inflation Forecasting

Typically, when using econometric techniques to forecast economic variables, estimation is carried out on a forecasting model that is built upon some assumed economic structure, based upon a priori knowledge and economic principles. However, such techniques cannot avoid running into the possibility of misspecification, which will occur should there be some error in the assumptions underlying this economic structure. Even when diagnostic tests have been easily cleared, a small change in the way this structure is set up can induce large differences in the forecast value. In other words, the researcher's subjective choices in setting up the model can have a substantial influence on the estimated forecast. In this paper, in which we concentrate upon inflation forecasting, we present a statistical forecasting method (SFM) that stresses statistical relationships among time series data, and that makes no structural assumptions, other than to set up the underlying variables. When putting together a forecast, this SFM first builds a number of VAR models from combinations of the underlying variables; it then automatically ranks these, based upon their performance. Furthermore, it has the additional property that it produces forecasts not merely by looking at the movements of the forecast themselves over time, but by taking into account the uncertainty in both the model and the forecast value captured in the forecast distribution (and illustrated in the fan charts). We also carry out analysis that looks just at the question of whether future inflation will move upwards or downwards, attempting to produce a qualitative forecast of this movement.