The Forecasting Performance of Dynamic Factor Models with Vintage Data

We present a comparative analysis of the forecasting performance of two dynamic factor models, the Stock and Watson (2002a, b) model and the Forni, Hallin, Lippi and Reichlin (2005) model, based on vintage data. Our dataset that contains 107 monthly US “first release” macroeconomic and financial vintage time series, spanning the 1996:12 to 2017:6 period with monthly periodicity, extracted from the Bloomberg database† . We compute real-time one-month-ahead forecasts with both models for four key macroeconomic variables: the month-on-month change in industrial production, the unemployment rate, the core consumer price index and the ISM Purchasing Managers’ Index. First, we find that both the Stock and Watson and the Forni, Hallin, Lippi and Reichlin models outperform simple autoregressions for industrial production, unemployment rate and consumer prices, but that only the first model does so for the PMI. Second, we find that neither models always outperform the other. While Forni, Hallin, Lippi and Reichlin’s beats Stock and Watson’s in forecasting industrial production and consumer prices, the opposite happens for the unemployment rate and the PMI.