Advisor Perspectives welcomes guest contributions. The views presented here do not necessarily represent those of Advisor Perspectives.
Recessions are notoriously difficult to forecast. That, of course, hasn’t stopped many high-profile analysts from predicting recessions in 2010 and 2011 – incorrectly, at least thus far. Given the wealth of often contradictory economic data that exists today on which to base such forecasts, this should come as little surprise. What’s more surprising, however, is that they have based their predictions on models that were ill conceived and insufficiently tested.
Among those pundits who have relied on faulty models are John Hussman, David Rosenberg and John Mauldin. I will examine some of the models they have used and then turn to a simple – but far more predictive – model that does not warn of a recession now.
Almost all recession forecasting is based on historical data and observing an indicator over time. A recession call occurs when certain patterns appear that have also preceded prior recessions.
I have developed an evaluation model that can assess how well various indicators have performed to date and assign each an effectiveness score. Appendix 1 has a detailed description of my scoring system, which uses official NBER recession dating to determine whether a call was correct or not. The maximum score is 1.00 for a perfect indicator. There is no downside limit for a bad one, but any score below zero raises serious questions about an indicator’s usefulness.
Armed with this assessment tool, let’s first look at a sampling of three recession indicators that have appeared in Advisor Perspectives commentaries.
The WLIg+5.7
The Economic Cycle Research Institute’s (ECRI) Weekly Leading Index (WLI) – or, more specifically, its growth figure (WLIg) – is a favorite indicator of many commentators. John Mauldin wrote about it on June 26, 2010, endorsing a view that when the WLIg falls as low as -5.7 – as had occurred at the time – it always, except for one occurrence, accurately signaled the past seven recessions.
Nothing could be further from the truth. From this assertion, one might expect the WLIg to be at or below -5.7% in the lead-up to each and every one of the past seven recessions. In fact, that level of the WLIg preceded only four NBER-dated recessions – those that began in 1970, 1973, 2001 and 2008. The WLIg never reached this level prior to or during the 1981 recession, and it got there only 9 and 12 weeks after the start of the 1980 and 1990 recessions, respectively.
What’s more, this indicator gave a false alarm in 1987 and at least one more since the beginning of 2010 – possibly a second, if a recession does not begin soon.
The chart below depicts this indicator’s history since 1968, with 5.7% added to the WLIg to place the trigger line at zero – hence, my shorthand for this indicator is WLIg+5.7. The blue line tracks the indicator’s value over time, while the shaded areas correspond to NBER-designated recessions.
This indicator WLIg+5.7 obtained a score of -2.331 from my evaluation model.
The WLIg13
Another indicator referenced in John Maudlin’s June 26, 2010 commentary was the 13-week annualized growth rate of the WLI, which I will refer to as WLIg13. At the time, this value was -23.46%, which Maudlin called “very weak.” All prior instances where this indicator fell that low reflected a recession that was either imminent or already underway, Maudlin asserted, and he used the following table to illustrate his point:
Mauldin credited the table and his interpretation of it to Chad Starliper of Rather & Kittrell.
When I crunched the numbers myself, I got slightly lower values, which I attribute simply to the fact that Starliper and I probably had slightly different calculation methodologies. For my data, then, I adjusted the trigger threshold in my own data accordingly, to -19.0%.
The data show that this indicator failed to recognize the 1981 recession – I found that its minimum value in November, 1981 was only -8.91%, way above my recession threshold. Further, it can hardly be claimed that this indicator convincingly alerted us to the 1990 and 2001 recessions, as it provided no useful advance warning. Its score was an even-worse -4.320 in my evaluative model.
The GDP-1.60
More recently, financial experts have issued many warnings of an impending recession, based in part on the low year-over-year growth in U.S. real GDP. This indicator was included in Maudlin’s July 30, 2011 commentary “An Economy at Stall Speed” and also in John Hussman’s August 1, 2011 commentary, “More Than Meets the Eye.”
Hussman included the following chart and used it to argue that the 1.6% real GDP growth prevailing at the time was slow enough to “anticipate an oncoming recession.”
For one thing, the premise of the above chart is not even quite accurate – the most recent data available from the Bureau of Economic Analysis at the time this article was published showed year-over-year GDP growth of 1.62%.
But the deeper problem with this graph is that it is not a real-time graph of this indicator, and thus incorrectly shows the indicator giving warnings of oncoming recessions. Chart 3 shows the recessions in real-time, while the indicator follows the GDP series dates, which are four months before the real-time dates when BEA releases the advance estimates of GDP for the relevant period being measured. The GDP series values are listed on the first day of the quarter to which they refer and are therefore four months displaced from real-time.
To plot a real-time graph for year-over-year growth in GDP, one has to add four months to the dates of the GDP series, because this is when one would for the first time have had some information about the magnitude of the latest GDP number. I have done this, and my real-time counterpart to the chart above is shown below.
The visual effect of bringing the data into real time is simply that the graph has shifted further to the right relative to the shaded recessions. For example, the most recent trough of the indicator now falls after the end of the last recession, no longer before the end of it as shown in chart 3. A significant difference if you are seeking advance – or even contemporaneous – warning of a recession!
This indicator, when plotted in real-time, is a lagging indicator; it signals the beginning of a recession on average 20 weeks late. Yet, when plotted following the GDP series dates, it is presented by many analysts – by no means Hussman and Maudlin alone – as a leading indicator that can somehow let us “anticipate an oncoming recession.”
The real-time GDPyoy-1.6 gets a very poor score of -4.239 from my evaluation model. One can see that this indicator was late to recognize recessions and indicated long periods of time to be in recession when there was none. Also it provided two false positive signals, extending over 26 weeks, between 2003 and 2007.
It’s not just a matter of calibration, either, though some could help. If we moved the trigger from 1.6% to 1.3%, the indicator would receive a better – but still not very good – score of -3.486, and one of the false positives would be eliminated. The same recession-indicating lag, however, would remain. This is simply not a fruitful approach.
Better indicators
The decisions to base these forecasts on the indicators examined above is doubly surprising when you realize that there are better indicators readily available. One is the six-month smoothed compound annualized growth rate of the Conference Board’s Leading Economic Index, which I will designate as CB-LEIg. Its real-time lag is about five weeks, because the index is published monthly.
Using the old series values to derive the six-month smoothed annualized growth rate, then adding 0.50% to its value to make zero the recession indicating trigger, one obtains a high score of 0.457 for this indicator. Its real-time chart is shown below.
This indicator did not warn of a recession in 2010 and 2011 and provided correct recession warnings for all of the previous seven recessions without any false alarms, an impressive track record. What’s more, its warnings were always coincided with or preceded the start of each recession as shown in the table below:
Recession |
Lead to recession start weeks |
1970 |
15 |
1973 |
9 |
1980 |
43 |
1981 |
0 |
1990 |
57 |
2001 |
16 |
2008 |
1 |
The CB-LEI is not some secret dataset to which only I have access. All the data above was available to the analysts whose recession calls were reviewed in this article, and they could easily have produced the same graph that I did.
Conclusion
The models used by Rosenberg, Maudlin, and Hussman have been poor predictors of oncoming recessions. Any reasonable person would use the Conference Board’s data in preference to any of the three indicators analyzed. The inaccuracies in both the original work of these analysts and the work they cited should have been detected; widely followed commentators have a responsibility to be accurate since they provide information to the public.
This state of affairs calls to mind the work of University of California-Berkeley psychologist Philip Tetlock, who has studied the forecasting of political pundits. “Human beings who spend their lives studying the state of the world … are poorer forecasters than dart-throwing monkeys,” a 2005 review of Tetlock’s work in the New Yorker concluded. “Knowing a little might make someone a more reliable forecaster, but … knowing a lot can actually make a person less reliable.”
It is not surprising that the experts’ dire 2010 recession predictions turned out to be incorrect, given the models they relied upon. But their 2011 forecasts cannot yet be discounted. Should a recession occur soon, these same experts will no doubt trumpet what is ultimately a dubious forecasting success; even a dart-throwing monkey hits the target every once in a while.
Appendix 1
Scoring System for evaluating Recession Capturing Indicators
Parameters needed to calculate a score:
(a) Total number of weeks in series.
(b) Total number of weeks in the seven NBER recessions from 1968 to 2011 (=360)
(c) Weeks captured before recession starts.
(d) Weeks captured prior to 26 weeks before recession starts.
(e) Weeks captured after recession ends, excluding any false positive and continuous string of positives to end of series.
(f) Weeks not captured after recession starts and before it ends.
(g) Weeks in continuous string of positives to end of series.
(h) Weeks with false positives excluding continuous string of positives to end of series.
(i) Number of instances when a false recession signal occurred. False signals separated by 7 weeks or less are counted together as one.
(j) Number of continuous positive weeks leading to recession start. This is a positive number if the first week of the lead occurred before the recession starts, and a negative number if the first week of the lead occurred after the recession starts.
(k) Number of continuous positive weeks leading to recession end. This is a positive number if the first week of the lead occurred before the recession ends, and a negative number if the first week of the lead occurred after the recession ends.
Constraints and penalties:
- The maximum allowable signal lead is 26 weeks before a recession start. Every week captured before this limit is subtracted from the total positive count. We do not want more than 26 weeks lead for a recession signal.
- Weeks captured after recession ends are doubled and subtracted from the total positive count. We do not want models with long tails.
- We really do not want false signals. The number of weeks of false positives is multiplied by (3 + nr. of false signals) and subtracted from the total positive count.
Early/Late Factor (ELF):
- Acceptable negative lead to the start of recession is 2 weeks. If the negative lead exceeds 2 weeks the models is penalized with a factor which takes the recession length into account.
- Acceptable maximum positive lead to the end of recession is 10 weeks. If the positive lead exceeds 10 weeks the models is penalized with a factor which takes the recession length into account. We are not too concerned to pin-point the end of the recession accurately.
- Acceptable maximum negative lead to the end of recession is 10 weeks. I the negative lead exceeds 10 weeks the models is penalized with a factor which takes the recession length into account.
Total number of weeks incorrectly captured used to calculate a score:
(f) Weeks not captured after recession starts and before it ends. (The ELF factor penalizes the model for this, thus we do not apply a multiplier to this number.)
(h) Weeks with false positives x (3 + nr. of false positives)
(d) Number of captured weeks prior to 26 weeks before recession start.
(e) Weeks captured after recession ends x 2
Score Calculation:
NBER capture
NBER = total number of captured weeks during recessions / 360
[360 – (f)] / 360
Area Under Curve
AUC = (total nr. of weeks in series – total nr. of weeks incorrectly captured) / total nr. of weeks in series
{(a) – [(f) + (h) + (d) + (e)]} / (a)
Model Score:
NBER* = (NBER – 0.90) x 10
AUC* = (AUC – 0.90) x 10 x ELF
Score = (NBER* + AUC*) / 2
The maximum possible score is 1.00
Appendix 2
Calculating the annualized growth rate of a time series
A superior measure for cyclical analysis, introduced by Geoffrey H. Moore, is the six-month smoothed compound annualized growth rate of a time series such as the WLI.
WLIg is the compound annualized growth rate of the WLI for a 26.5 week period. MA1 is the 4 week moving average of the WLI and MA2 is the moving average of MA1 over the preceding 52 weeks. Because the 52-week average in the denominator is centered 26.5 weeks before the current middle of the week, the ratio MA1/MA2 yields the change over a 26.5-week period, i.e. over six months.
WLIg = [100*(MA1/MA2)^( 52/26.5)] – 100
One can also calculate the WLIg using the Excel function XIRR.
=XIRR(cell 1: cell 2, (date)six-months-ago : (date)now)
where cell 1 contains: -1 and cell 2 contains: MA1/MA2now
MA1 is not limited to 4 weeks, however this number is used for the WLI.
WLIg13 is the average of WLIg12.5 and WLIg13.5.
WLIg12.5 is the compound annualized growth rate of the WLI for a 12.5 week period. MA1 is the WLI for the current week and MA2 is the moving average of MA1 over the preceding 24 weeks. Because the 24-week average in the denominator is centered 12.5 weeks before the current middle of the week, the ratio MA1/MA2 yields the change over a 12.5-week period.
WLIg13.5 is the compound annualized growth rate of the WLI for a 13.5 week period. MA1 is the WLI for the current week and MA2 is the moving average of MA1 over the preceding 26 weeks. Because the 26-week average in the denominator is centered 13.5 weeks before the current middle of the week, the ratio MA1/MA2 yields the change over a 13.5-week period.
One can calculate the WLIg12.5 and WLIg13.5 using the Excel function XIRR similar to what is shown for WLIg above.
Georg Vrba is a professional engineer who has been a consulting engineer for many years. In his opinion, mathematical models provide better guidance to market direction than financial “experts.” He has developed financial models for the stock market, the bond market and the yield curve, all published in Advisor Perspectives. The models are updated weekly. If you are interested to receive theses updates at no cost, or to have a recession indicator series scored according to the evaluation model described in this article, send email request to .
Read more articles by Georg Vrba, P.E.
