On the advantages of disaggregated data: Insights from forecasting the US economy in a data-rich environment

The good forecasting performance of factor models has been well documented in the literature. While many studies focus on a very limited set of variables (typically GDP and inflation), this study evaluates forecasting performance at disaggregated levels to examine the source of the improved forecasting accuracy, relative to a simple autoregressive model. We use the latest revision of over 100 U.S. time series over the period 1974-2009 (monthly and quarterly data). We employ restrictions derived from national accounting identities to derive jointly consistent forecasts for the different components of U.S. GDP. In line with previous studies, we find that our factor model yields vastly improved forecasts for U.S. GDP, relative to simple autoregressive benchmark models, but we also conclude that the gains in terms of forecasting accuracy differ substantially between GDP components. As a rule of thumb, the largest improvements in terms of forecasting accuracy are found for relatively more volatile series, with the greatest gains coming from improvements of the forecasts for investment and trade. Consumption forecasts, in contrast, perform only marginally better than a simple AR benchmark model. In addition, we show that for most GDP components, an unrestricted, direct forecast outperforms forecasts subject to national accounting identity restrictions. In contrast, GDP itself is best forecasted as the sum of individual forecasts for GDP components, but the improvement over a direct, unconstrained factor forecast is small.