Disaggregated prediction system - Forecasting evaluation for inflation 2006-2011

The study describes the characteristics of the Disaggregated Prediction System (SPD) within the literature’s boundaries and presents the evaluation of its forecasting performance in terms of the Lima’s CPI 12-month percentage changes, the core inflation index and its complementary non-core inflation index during the forecasting period July 2006 – May 2011. This ex post evaluation considers not only the different performances of the main multivariate specifications included in SPD but also the auto-regressive univariate benchmark model for each aforementioned sub-aggregate. The evaluation considers two root-mean-squared-forecasting error (RMSFE) computations: (i) the static evaluation uses the whole sample of prediction errors selected for each horizon h, and (ii) the dynamic version uses the prediction errors inside fixed-width moving windows (whose upper limit is period ?) selected for a pre-defined horizon h. In line with the literature, the ex post static evaluation shows the convenience of forecasting the disaggregated components with multivariate models, in contrast with using the sub-aggregate series and forecast it with the univariate models. The main result obtained from the dynamic evaluation is the presence of "crossings" in the corresponding temporal evolutions of the RMSFEs for many groups of models: the previously good performance of a group of models becomes a not-so-good performance (Aiolfi & Timmermann(2006)). These "crossings" justify using "combined forecasts".