Revision Risk in Real-Time Macroeconomic Forecasting

Abstract

Macroeconomic forecasts refer to outcomes that are first released and then revised. A 90 percent interval for the first GDP release, a six-month value, or a latest-value benchmark is not the same uncertainty statement. We ask how revision risk evolves through the release cycle and what can be reported in real time when later-outcome errors are scarce. We decompose later-outcome MSE into preliminary forecast risk, revision risk, and their covariance. In SPF data, first-release to roughly 180-day revisions account for 8.3 percent of later-outcome MSE across real-activity targets, versus 3.6 percent across inflation targets. We show that later-outcome uncertainty is partially identified: released histories give early-error and revision marginals, but not their dependence. This yields a sharp Frechet-Makarov set and motivates direct late calibration, dependence-robust transport, and signed or revision-model transport. Out-of-sample results support method choice rather than a universal transport rule: coverage and stability determine when transport gains are usable.