This paper explores the role of model and vintage combination in forecasting, with a novel approach that exploits the information contained in the revision history of a given variable. We analyse the forecast performance of eleven widely used models to predict inflation and GDP growth, in the three dimensions of accuracy, uncertainty and stability by using the real-time data set for macroeconomists developed at the Federal Reserve Bank of Philadelphia. Instead of following the common practice of investigating only therelationship between first available and fully revised data, we analyse the entire revision history for each variable and extract a signal from the entire distribution of vintages of a given variable to improve forecast accuracy and precision. The novelty of our study relies on the interpretation of the vintages of a real time data base as related realizations or units of a panel data set. The results suggest that imposing appropriate weights on competing models of inflation forecasts and output growth — reflecting the relative ability each model has over different sub-sample periods — substantially increases the forecast performance. More interestingly, our results indicate that augmenting the information set with a signal extracted from all available vintages of time-series consistently leads to a substantial improvement in forecast accuracy, precision and stability.
data and model uncertainty forecast combination real-time data