Nowcasting U.S. GDP with a Mixed‑Frequency Dynamic Factor Model with Student‑t Errors

We ask whether heavy-tailed innovations improve the nowcast of U.S. real GDP growth. A mixed-frequency dynamic factor model with Student-t innovations, so that each observation reaches the state weighted by its own surprise, is estimated by expectation maximisation and scored week by week against four Bayesian VARs, the New York Fed Staff Nowcast and two univariate benchmarks, on one panel of thirty-seven series in a pseudo-real-time exercise over 2007–2025. The tails are in the panel beyond doubt: in sample the Gaussian restriction is rejected by a margin no critical value comes near. The nowcast gain is far smaller. Before the pandemic the Student-t leads the Gaussian, significantly in the deep forecast and in the backcast, and the lead is gone once the pandemic quarters enter. The factor model carries a lower average error than all four Bayesian VARs and than the New York Fed, yet each Bayesian VAR captures more of the four extreme quarters, on average, than any factor variant. We conclude that robustness to extreme observations is not robustness to extreme events.

The panel below replays the 2008–09 crisis as the model lived it: every Friday from October 2007 to January 2010, each quarter's nowcast is updated with the data released that week, one colour per quarter, until the BEA prints the real number (the dot). First with Gaussian errors, then with Student-t errors on the same calendar. The fat tails stop a single extreme release from dragging the whole path, and the gain shows in 2009, as the economy turns.