The Latest Jobs Report And Recession Probability Models

The December Nonfarm Payrolls report was much stronger than expected, which reduced recession worries. I will leave the dissection of the report internals to others, but I just want to point out how these data fit in with my current research topic: recession probability models. It is safe to say that there is a divergence between activity data and sentiment indicators (including market pricing, such as the yield curve). Recession models based on activity data have one advantage in the current environment: they are unlikely to register a false positive.

I will keep this comment fairly brief, as it effectively is a spoiler for some upcoming articles (which are preliminary drafts for a handbook on recession analysis). What I am interested in are models that give an explicit probability forecast of being in recession. This is in distinction to models that generate an indicator that rises and falls in line with economic activity, such as leading indicators, or principal component (PCA) models. We need to come up with some rule to convert the indicator reading into a recession probability.

Building recession (as well as related concepts, such as financial stress) probability models is currently a research fad. Unsurprisingly, they use all kinds of novel econometric techniques to develop the indicator. The academic research I have seen emphasises the mathematical structure, but I would divide them into two categories.

1. Models that create a recession probability based on the trajectory of real economic variables. The simplest (and oldest) was to base this on real GDP, but later models looked at the simultaneous developments in key monthly data. (In particular, the economic series emphasised by the NBER in their recession dating methodology.) Furthermore, it is possible to look at a cross section of regional data (e.g., employment at the state level in the United States).
2. Models that attempt to forecast recessions by using leading indicators, the most popular being some yield curve slope.

The first class of models have an obvious limitation: they are essentially coincident indicators, and cannot truly be used to forecast a recession. At most, they can be used to give a recession call before the official pronouncement by the NBER. That said, that can be quite useful: markets and market strategists were in denial about the U.S. being in recession long after December 2007.

Their advantage is the general lack of false positives. Under normal circumstances, they seem to be equivalent to the NBER recession-dating methodology, except that they are purely quantitative models. (There is no guarantee that this is the case, however.) It allows the decision-making process to get away from duelling chart packs, which is a common feature of economic turning points.

Leading indicator-based models attempt to look ahead, which theoretically makes them more useful. The problem is straightforward: they are calibrated by looking at events that have been happening once per decade in the United States over the post-1990 period. I am naturally suspicious of mathematical models; I'd only be happy with the out-of-sample testing of such models by around 2060 or so.

At present, we can see the divergence in models. Activity-based models for the United States will generally be chugging along with a 0% recession probability (although I guess one could create one that has somehow triggered), while models that are based on market indicators may have had their recession probabilities shoot up in recent months. (Which is the sort of thing you need to get used to if you work with mathematical economic models.)

The emphasis on the yield curve or credit spreads in forecasting recession probability models makes them highly suspect for fixed income investors. You are supposed to be valuing the level of the yield curve based on your probability forecasts. (For readers who are not familiar with fixed income pricing, a recession is normally associated with hefty policy rate cuts. Bond yields are based on the expected average of policy rates over the bond's life span; if rates are expected to fall, longer-term bonds have a lower expected average yield. This means that the curve ends up inverted -- longer-term bond yields are below shorter-term yields.)

If your recession odds are based on the yield curve, you have created a circular information loop in your investing process: the more expensive the long end gets versus the short end, the more attractive it is fundamentally. The result is that you end up being a trend follower: you buy bonds when yields are falling (price up, yield down!), and sell when they are rising. Trend following strategies tend to work in other, lesser, markets, but tend to be suicidal in fixed income, where range-trading works 90% of the time. (That other 10% is when trend following crushes range-trading.)

Even if you are not a fixed income investor, you should probably question whether fixed income pricing is always going to be reliable going forward as it has been historically. The largest problem we faced in recent years was the presence of the zero bound; even though negative interest rates are possible, it is very hard to get them very negative. This creates a positive bias to the curve. The next problem is that it is possible that something has gone wrong in fixed income positioning (fixed income markets see a lot of leveraged trading) -- or that fixed income investors are just plain out-to-lunch with regards to economic prospects.

I expect to return to the question of the activity-based recession probability models in my next article.

(c) Brian Romanchuk 2019