gentlemen. Steven Kopitz wrote:
But in general, the 12mm-based VMT will turn a little bit to the right before the recession starts.
I take a non-seasonally adjusted series of vehicle mileage going back to 1970 and define a 12-month moving average variable. I think of defining a dummy variable that takes a value of 1 whenever this moving average term becomes negative. This is plotted as blue bars in the graph below, with NBER-defined decay dates in dark grey:
figure 1: Months in which the VMT’s 12-month moving average declined (blue bars), and NBER-defined recession dates (dark gray bars). Source: DOT FHA through FRED, NBER and author’s calculations.
Using a lagged probability model, with a negative 12-month VMT virtual moving average as a predictor, I therefore predict a recession:
figure 1: Implied recession probabilities using dummy variables with uniform values when the 12-month trailing moving average of vehicle miles traveled is negative, one month lag (pink), and NBER-defined peak-to-valley recession dates (grey bars). Source: NBER and author’s calculations.
The McFadden R2 for this regression is 0.11. 48% correctly predicted recessions (using a 30% threshold to predict recessions). Regression output at the end of the post.
Clearly, there is a lot of misjudgment, precisely because in many cases the 12-month moving average of the VMT declines without accompanying a recession as defined by the NBER (of course, one can define a recession in a way that makes the VMT variable perfectly predictable) ); that would define a recession as a few months after the 12-month moving average of the VMT fell! )
What if we used a 12-month tracking moving average (lag 1 month) to actually change? Then we get this picture.
image 3: Recession probability is implied using a 12-month trailing moving average of vehicle miles traveled, a one-month lag (sky blue), and NBER-defined peak-to-valley recession dates (grey bars). Source: NBER and author’s calculations.
The McFadden R2 for this regression is 0.04. 16% correctly identified recession (using the 30% threshold again). Regression output at the end of the post.
Again, this is not a particularly good predictor of a recession as defined by the NBER.
appendix:
The VMT variable is a reasonable coincidence indicator. The McFadden R2 for probabilistic regression was 7%, with an 18% correct call decay using a threshold of 30%.
Figure 4: Recession probability is implied using a 12-month trailing moving average of vehicle miles driven over the same period (green) and NBER-defined peak-to-valley recession dates (grey bars). Source: NBER and author’s calculations.
