Ever wonder whether vehicle miles traveled (VMT) does a good job of predicting recessions. Wonder no more. First take a look at what VMT does over recessions, versus heavy truck sales (suggested by Calculated Risk at some points), and the eponymous Sahm Rule (real time version).
Figure 1: 12 month growth rate in the vehicle miles traveled, n.s.a. (teal), in heavy truck sales, s.a. (tan), and Sahm rule indicator – real time (black). Sahm rule is 3 month moving average unemployment rate relative to lowest unemployment rate in last 12 months. Orange dashed liine denotes threshold for Sahm rule indicator. NBER defined peak-to-trough recession dates shaded gray. Source: FHA via FRED, Census via FRED, FRED, and NBER.
It’s hard to see, but the 12 month change in VMT declined a few months ago (the level is lower than pre-pandemic), while heavy truck sales are up, y/y. The Sahm rule is near zero (it needs 0.5 ppts to breach the threshold).
How do these variables do in predicting recessions (peak-to-trough, as defined by NBER BCDC, not the amorphous definition used by certain individuals)? Here’re the probit regression results over the 1970M01-2022M10/M11 period (where I convert the Sahm rule variable to a dummy variable (Sahmruleindex) taking on a value of 1 if the Sahm rule variable exceeds the 0.5 threshold).
Notice that the y/y growth rate of VMT doesn’t explain a lot of recessions, according to the McFadden R2. The y/y growth rate of heavy truck sales actually explains the most. The Sahm rule index has an intermediate explanatory power of 18%. The Sahm rule is supposed to signal a recession has started, not necessarily a recession is currently ongoing.
Below I plot the predicted probabilities associated with the estimated VMT and heavy truck sales probits, and the implication of the Sahm rule (nonestimated).
Figure 2: Probit estimated recession probabilities from 12 month growth rate in VMT (teal), in heavy truck sales (tan), and implication from interpretation of Sahm rule (black). NBER defined peak-to-trough recession dates shaded gray. Source: NBER, and author’s calculations.
Since it’s hard to see what’s going on, particularly in recent recessions, I focus in on the period starting just before the Great Recession (corresponding to Figure 1).
Figure 3: Probit estimated recession probabilities from 12 month growth rate in VMT (teal), in heavy truck sales (tan), and implication from interpretation of Sahm rule (black) (detail). NBER defined peak-to-trough recession dates shaded gray. Source: NBER, and author’s calculations.
Now, VMT would’ve done better if I truncated the sample at 2019M12, given the extreme nature of VMT behavior during the pandemic. However, the McFadden R2 only rises to 12%, and still wouldn’t have breached the 50% threshold for the 2007-09 recession (but would’ve predicted a recession in 1996).
…so you’re saying there were 1.1 m more jobs and gasoline consumption and VMT were falling at the same time? So not only did those extra 1.1 m workers not drive to work, those with jobs were also driving less. It’s possible, sure. But if I have to adjudicate between the CES and the HH survey, the data is more consistent with the HH survey. But that’s not what Menzie did. And neither did you. But I did, and as it turned out, that inference appears to have been correct.
My point is that VMT is particularly unreliable from a formal standpoint, and it is reasonable to presume a structural break in the employent-VMT relationship given recent developments, including working from home.
To illustrate the potential break in this employment-VMT relationship, I convert to quarterly data the monthly, and run a regression in first log differences. A recursive residuals one-step-ahead Chow test rejects stability around the pandemic.
Figure 4: Recursive one-step ahead residuals from indicated regression (blue, right scale), 95% interval (red dashed, right scale), probability of no change (blue circle, left scale). NBER defined peak-to-trough recession dates shaded gray. Source: author’s calculations using EViews.
- One shouldn’t rely on VMT to infer employment given the structural breaks identified in the data.
- There is little indication that as of November 2022, or in 2022H1, that the economy was in recession (as defined by the NBER Business Cycle Dating Committee).
Of course, random observers can define a recession any way they want, so as to make their declarations valid. I.e., one could call a recession one where the Michigan index falls below a certain value. But that would be outside of the spirit of the business cycle literature.
Addendum, 2pm Pacific:
Mr. Kopits also suggested use of gasoline use. I also estimated a probit regression using 12 month growth rate in gasoline supplied (EIA). While the coefficient on gasoline is statistically significant, the McFadden R2 is low (0.04), and fails to predict any recession save the 2020 one (using a 50% threshold).