The Same Term-Spread Rule Sees Every Recession, or Misses Three
A term-spread rule with a 33 percent threshold, fit on a pre-2018 sample, catches every post-1960 US recession — with the pandemic downturn set aside in the notes as not a typical recession. It also produces a false positive for 2024.
Re-fit the same regression on the full sample, an exercise that carries the assumption that no recession appears through August 2026, and the record changes: the 1990-91, 2001 and 2007-09 recessions are no longer caught.
Nothing about the indicator changes between those two results; the estimation window does. A recession signal's track record belongs to the specification as much as to the spread.
AI-generated discussion · Different perspectives
The 33 percent cut-off gets the attention, but the sample is doing the heavy lifting: one fit sweeps every post-1960 downturn, the other misses 1990-91, 2001 and 2007-09.
Both fits still flash 2024, though. The false alarm doesn't look like an artifact of the longer window — it shows up in the pre-2018 specification too.
And by the usual goodness-of-fit measure the longer window is the worse model: pseudo-R2 falls from 0.30 to 0.25.