Consumer confidence poorly predicts recessions.

The figure below combines Blanchflower and Bryson’s Chart 2 [of Sentiment Indices] (on top) with my plot (over the same sample period) of the 10yr-3mo term spread and spread adjusted by the NY Fed’s 10-year term premium (at bottom).

Figure 1: Top -Blanchflower and Bryson (2021) Chart 2, of Univ. of Mich. Expectations Index (red dashed line) and Conference Board Sentiment Index (green line); Bottom – 10yr-3mo US Treasury term spread, % (blue line), and term spread adjusted by estimated term premium (brown line); NBER defined recession dates shaded gray. Source: Blanchflower and Bryson (2021), and Federal Reserve Board, NBER, and author’s calculations.

There is indeed a sharp drop in the two sentiment indexes. On the other hand, the term spread does not signal an incipient recession. To formalize this point, let’s examine the predictions of each indicator, individually and combined, using a 12-month horizon, over the 1978-2019 period (so excludes recession information for 2020). I only test the Michigan sentiment index as I don’t have access to the time series for the Conference Board measure.

P(recessiont+12=1) = 0.11 – 0.015 sentimentt

McFadden R2 = 0.02

P(recessiont+12=1) = -0.317 – 0.693 spreadt

McFadden R2 = 0.29

P(recessiont+12=1) = 1.125 – 0.686 spreadt – 0.017 sentimentt

McFadden R2 = 0.31

Bold face denotes statistical significance at 5% msl.

On its own, sentiment doesn’t show up as explaining a lot of the movements in the recession indicator, perhaps due to the fixed lag length of 12 months. The conventional 10yr-3mo Treasury spread explains a lot more of the recession indicator; the incremental information added by the sentiment indicator is small.

For predictions of recessions, one can see the comparative performance in Figure 2.

Figure 2: Probability of recession predicted from Michigan sentiment index (blue), from 10yr-3mo Treasury spread (brown), and from sentiment index and Treasury spread (teal). Red line at 50% threshold. NBER defined recession dates shaded gray. Source: Univ. Michigan, Federal Reserve via FRED, NBER, and author’s calculations.

The Michigan sentiment indicator certainly covaries and precedes recession dates, although the probit model does not do a particularly good job at detecting big changes in probability (average is about 10%, rising to about 20% at actual recession times).

Note that Blanchflower and Bryson focus in on big changes of 10 percentage points are more, pointing to a nonlinearity I’ve not incorporated (although I did try changes in sentiment as an indicator, yielding similarly lackluster results.)

Note that the term spread does a better job, but not altogether great. The candidate reasons for this result are manifold, as discussed most recently in this post. Using both variables improves the fit slightly, but does not yield a specification that is appreciably better at predicting recessions.

As for whether we’ve already entered into recession, I’m agnostic. My recollection of a discussion between myself and another CEA person in April 2001 was that we were assuredly not in a recession at that time. Of course, we only had data up to February, maybe a bit of March; and all that data was subsequently revised.

NBER eventually dated the peak at March 2001. However, to the extent we have high frequency indicators we didn’t have back then, it doesn’t appear we were in a recession as of early-mid October.