Through a combination of quackery, charlatanism, and inadequate utilisation of mathematics, callers for US recession in 2016 are embarrassing themselves. Again.
The most prominent reason for recession calling may well be the Institute of Supply Management’s Manufacturing Purchasing Manager Index. The problem with this recession forecasting methodology is that it doesn’t work.
The first problem is statistical. The squared correlation coefficient between recessionary state and ISM PMI is 0.33 (33%) with a standard error of 0.29 (29%). This means that, when you calculate the recession probabilities from ISM PMI, they are only good within +/- 29%. Ouch.
The second problem is that of false positives. We can easily eyeball nine prior examples of the present implied recession probability being matched or eclipsed, but generating no recession. Ergo, using ISM PMI to forecast a recession is a Type I error.
Finally, when the standard error of the model (29%) is most of the useful threshold for determining recessionary state (43%). And, of course, the last reading of the model (26.5%) is less than the standard error.
If anyone tries to talk to you about recession probabilities using ISM PMI as data to back their argument, they’re doing it wrong.
How is any of this even possible, you ask? The answer is that manufacturing is not a material relative to the size of the balance of the US economy. Evidence of this is observed by regressing Real GDP against manufacturing and service ISMs.
The coefficient for service ISM is over 36 times that of the manufacturing ISM. In fact, if you want to forecast recessions using ISM data, non-manufacturing ISM gets you much closer.
The squared correlation coefficient is 0.48 (48%) vs 0.33 for manufacturing, and the standard error is 0.23 vs 0.29 for manufacturing. The useful threshold for determining recessionary state (54%) is 2.3x the standard error, instead of just 1.48x for manufacturing ISM. Unfortunately, the data only starts in the late 1990s, so we are unable to make a meaningful comparison of false positives.
Oh, and the last reading of recession probability from non-manufacturing ISM is 5.4%.
If we’re going to use just a single variable for recession forecasting, then we should use a high signal, leading variable. The Leading Index for the United States, which is an amalgamation of state leading indexes, which themselves are a compilation of individual state-level economic indicators, is a good bet. Of course, it’s probably cheating a little methodologically, because it already is composed of many variables, but it makes for easy econometric work. Let’s try it.
With a squared correlation coefficient of 0.53, it’s the best fit of them all. The standard error, 0.21, is very small relative to the recession threshold of 59%. The last reading is a 6% probability.
If we’re serious, however, we won’t just use one variable. Univariate analysis makes for good charts and slides, but it’s a poor analytical method. It is also far more useful to look at recessionary probabilities simultaneously at different forecasting intervals. Here is a probability estimate using a multivariate model of my favourite leading indicators across 5 time-frames.