(Note: this is an unedited excerpt from my upcoming book on recessions. It is in a chapter that discusses empirical recession indicators. Since the concept was already introduced, the discussion here does not cover background concepts, such as what a recession indicator is used for.)
The rule is straightforward: the 3-month average of the (U-3) unemployment rate needs to rise by at least 0.5% versus its values over the previous 12 months, not including the current month. I used a similar construct in an earlier article (link), but I included the current month to floor the value at zero. If one wanted to have the clean-looking floor but replicate Sahm’s numerical results, just compare the 3-month average to the minimum over the past 13 months, including the current month.
The figure above shows the results of the indicator for the interval 1970-2018 (the time frame looked at by Sahm). As can be seen, it captures the NBER-defined recessions, with no false positives.
Rule RobustnessOne immediate concern would be the broader applicability of the rule. When we apply the rule to Canadian data (below), we see some discrepancies between it and the C.D. Howe’s Business Cycle Committee’s recession dating decisions.
as was noted in my earlier article, the Bank of Canada did cut rates during recent slowdowns that were not defined as recessions. As such, one could argue that the trigger rule has some advantages over the methodology used by the C.D. Howe committee. (I am agnostic about this question; as I will discuss below, I think the trigger rule is more useful for my purposes.)
The next angle of discussion is the construction of the rule. In my experience, some analysts place too much emphasis on the exact details of indicator construction. That is, we need to replicate the rule exactly, and apply the rules as specified in the original paper. Any deviation from the original result is viewed as something completely different. This perhaps reflects the culture of economics academia for defining original research. However, we need to be flexible and realise that that there is an infinite number of time series filters that will end up get essentially identical performance on a historical data set, but will end up giving different trigger dates in forward-looking data. Since we are not working with models of economic fundamentals, there is no theoretical reason to prefer any particular variant over the other; what matters is that they are capturing the same underlying dynamics.
Although the difference does not appear to matter much when looking at back history, it affects how the indicator is used in real time. What happens when the indictors rise close to the trigger levels that were determined by analysis of the back history? For example, the deviation from trend indicator might hit its trigger level in a month, while the indicator using Sahm’s construction might be short of its trigger. If we want to use a recession indicator as a binary on/off signal input to investment decisions (for example), this could easily affect performance (since returns are quite volatile around turning points). If we are back-testing the indicators, there are likely to be only a few points in time when they deviate from each other, and so the evaluation is highly dependent on one or two episodes.
In order to avoid this sensitivity to a few episodes, we need to make performance comparisons in some probabilistic fashion, which provides a smoother signal than relying on the small sample of indicator transition dates. However, a smooth evaluation means that qualitatively similar indicators – like the trigger series and the deviation from trend – will end up evaluating similarly. Although I feel that is a sensible outcome, there is a tendency for some people to demand the “best” or “optimal” indicator – which then leads to over-confidence in the indicator chosen.
Unemployment Rise Triggers and RecessionsThe success of this rate trigger indicator in the United States aligns with my personal bias towards such an indicator. My argument is straightforward: under current institution conditions, what matters about a recession is a rapid rise in unemployment. (I discuss institutional change issues below.) From a practical matter, I would be tempted to define a recession as a rapid rise in unemployment, and if people wish to have recession dates, use such an indicator to determine beginning and end dates for recessions. The implication is that the unemployment rate indicator would have a perfect recession calling performance (by definition).
From a theoretical perspective, one could argue that the unemployment rate has some deficiencies in this regard; we should be looking at employment growth. For example, it is theoretically possible that the unemployment rate could rise during an expansion if people who gave up on looking for jobs suddenly decided that things have improved, and they all started looking for jobs at the same time. (In order to count as unemployed, one needs to have tried to look for work, with the exact definitions used being determined by the surveys sent to polled households by national statistical agencies.)
Although such issues will cause a certain drift in the unemployment rate during an expansion, the slow-paced nature of these attitude changes means that they are not enough to cause a rapid rise in the unemployment rate. (The evidence for the previous assertion is the fact that the recession trigger has been a reliable recession indicator historically.)
The advantage of the unemployment rate over employment measures is that they are more stable over time. The absolute size of the working age population changes over time, so we need a way to convert the number of people working to be independent of the size of net number of jobs created. This means that we effectively have to look at something like a employment ratio, or divide through by the working age population – which looks the same as a ratio when the working age population changes relatively slowly (which has been the case in recent decades). Admittedly, the markets obsessively focus on the Nonfarm Payrolls monthly employment change number, but the author argues that this reflects of suitability of that statistic for gambling.[ii]
The bottom panel shows an attempt to build a trigger indicator based on the employment ratio, but with the sign flipped. (I took the maximum over 13 months including the current month, to get a nice clean upper bound at 0%.) Interestingly, the indicator does not appear to work as well as the unemployment rate-based indicator.
Structural Change RiskThe disadvantage with just using one time series as an input to the determination of a recession is that it is vulnerable to structural changes that affect that series.
The most extreme example of a structural change would be a switch over to a Job Guarantee programme (as favoured by Modern Monetary Theory). If the programme were implemented as intended, the “unemployment rate” might considered to be 0%, since everyone who is able to work would have a job that is available.[iv] (Realistically, many people such as professionals would stay outside the Job Guarantee and instead search for jobs matching their skills. They would not be involuntarily unemployed, whether they would be part of the measured unemployment rate is unclear.) We might need to switch over to an indicator based on the percentage of the working force within the Job Guarantee programme; a rise in this percentage would indicate a contraction of private sector employment. Calibrating such an indicator would be difficult initially, as there would be no past history to work with.
Less extreme changes could also affect the measured unemployment rate. For example, a change to the methodology could change the reported unemployment rate, which either disguises a contraction in activity, or create a spurious rise in the unemployment. Although this event could be compensated for by users of economic indicators, it is extremely awkward for a hypothetical recession definition that is only based on the unemployment rate. For this reason, I would not push very hard for defining recessions based solely on the unemployment – even though it may be the best definition for periods with no such structural changes.
Concluding RemarksThe unemployment rate indicator proposed by Claudia Sahm looks reasonable, and probably captures exactly what we are looking for when discussing recession risks. One can modify the rule used and get qualitatively similar results, so it should not be viewed as the last word on such indicators. The other leg of Sahm’s research – its use as part of a fiscal policy rule – which I will return to later.
[ii] Pretty well every other statistic is presented rounded off to the nearest 0.1%. The Nonfarm Payrolls number is presented in terms of thousands of jobs, and typically range between -100,000 to +200,000 jobs on the month. The illusionary precision means that betting pools or forecaster accuracy exercises will not have many ties, which would be the case for statistics that are rounded off to the nearest 0.1%.
[iii] The significance of the aging of the population as an explanation for the slow recovery of the employment-to-population ratio was a hotly debated topic in the aftermath of the Finance Crisis. The author was of the opinion that the demographic effects were greatly overstated, but that is not germane to the discussion at hand.
[iv] People who are acting in a disruptive manner would be ejected from the programme under most proposals; so unless they find a private sector job (presumably unlikely), they would be unemployed. The hope is that such people would not be a large percentage of the labour force.
(c) Brian Romanchuk 2019