Michael Dueker is a senior portfolio strategist at Russell Investments and formerly was an assistant vice president in the Research Department at the Federal Reserve Bank of St. Louis. He has been doing some very interesting economic research recently in developing what he calls a Qual VAR model for predicting recessions. We are pleased that he agreed to share some of the current implications of that research with Econbrowser readers, subject to the disclaimer that the content is the responsibility of the
author and does not represent official positions of Russell Investments
and does not constitute investment advice.
Current recession forecasts indicate that it would be unprecedented not to have a recession in 2008
‘recession’ contract, which is based on two consecutive quarters of
negative GDP growth, the two-quarter rule rarely holds at the time the National Bureau of Economic Research
(NBER) business cycle dating
committee declares business cycle peaks and troughs. At first glance, it might appear that
there is a close relationship between recessions and GDP growth
because of the well-known rule of thumb that a recession corresponds to two consecutive negative
quarters of GDP growth. It turns out, however, that this rule only works well for GDP data that
are revised well after the fact. For real-time GDP data, the two-quarter rule is not at all an
accurate recession guideline. Real-time vintage GDP data from the St. Louis Fed’s
Alfred database show that, among
the last eight business cycle turning points declared by the NBER,
only the peak and trough associated with the 1990-91 recession conform
to the two-quarter rule on the dates that the NBER announced the turning points.
In recent months, many economic
commentators, such as former Federal Reserve Chairman Alan Greenspan, have offered views on the
of a U.S. recession
in the near term. A large percentage of these assessments, while useful, are
judgmental in nature and ought to be compared with formal, model-based
Here I present current out-of-sample recession probabilities and prior forecast results.
One feature of recession forecasting is that statistical models rarely give
ex ante recession probabilities above 50 percent.
Birchenhall et al. (2001) serves
as an illustration of the way that modelers convert model-based recession probabilities
into yes/no recession calls at a probability threshold well below 50 percent. Of course, this
probability threshold will depend somewhat on the forecast horizon because longer-run forecasts
are hard-pressed to rise much above the unconditional probability of recession, which is about
17 percent. For the model considered here, the 2001 recession provides some guide as to
the probability threshold that one might apply for different forecast horizons.
In order to compare with judgmental recession odds, we have to think about
what Alan Greenspan means when he says that the odds are even that a recession
will occur in 2008. He means that if you take his words as a recession call, there
is at most a 50 percent chance that the call will turn out to be a false alarm.
Accordingly, one way to think about comparing a model-implied probability with a threshold is
to ask how likely it is that a probability above the threshold will turn out to be
a false alarm. For this reason, recession-forecasting models might apply a 32 percent threshold
and claim that there is a 70 percent chance that a recession call based on this probability
threshold will not be a false alarm.
Now the latter claim is hard to prove due to the small number of recessions observed in any
sample, but this translation puts judgmental recession calls and a model-based call on the same
The specific forecasting model I use is a Qual VAR from
It estimates and adds a business cycle index to a conventional vector autoregression (VAR)
of macroeconomic data.
The observable macro data included in the model are GDP growth, core CPI inflation, the slope of
the yield curve and the federal funds rate. (Note that the recession prediction model
examined here should not be confused with recession recognition models, such as
Chauvet and Hamilton (2006). The latter models discern recessions as soon as they
are evident in
the data without trying to look to the future.) Numerous studies have shown the slope of the
yield curve to be a powerful predictor of the business cycle, especially when combined with
information regarding the level of the real federal funds rate. The model is estimated using
Bayesian methods and the sign of the business cycle index is designed to conform
loosely with the NBER recession classifications, such that the index is generally
negative during recessions and positive during expansions.
One particularly appealing property of the Qual VAR is that because the business cycle index
arises from an equation in a VAR, dynamic forecasts of it are easy to compute.
In order to assess a recession forecasting model, it is important
to backtest it with real-time vintage data to avoid using data that were not available
at the time. Because the model makes use of the NBER recession classifications, it is
important not to report results of forecasts made at times when it was not clear how
the NBER would ultimately date a concurrent or recent recession. When trying to look ahead from
an economic expansion to
predict the onset of a recession, the forecaster has a pretty clear idea of whether the
most recent data come from a time period whose ultimate NBER classification is in doubt.
For example, in the current context in forecasting from the fourth quarter of 2007, we can
be reasonably certain that the NBER will not classify 2007Q4 as a recessionary one
and we can condition out-of-sample forecasts accordingly.
For this reason, the Qual VAR business cycle index model is estimated for the most recent 49
vintages of U.S. GDP data, which gives a set of real-time forecasts back to 1995Q3.
For each vintage, we know the precise history of GDP data that was
available at the time.
Two-quarter-ahead forecasts are shown in the first chart.
The two-quarter horizon allows us to examine how the model
predicted the first two of the three recession quarters in 2001.
Here one might assign a threshold probability
level of 35 percent, in which case the model successfully called the 2001 recession ahead of time
and it also is sounding an alarm at present. On the downside, using this metric, the model was
sounding a false alarm regarding the second half of 2007, based on data available six months
earlier. It might have been the case that the yield curve was signalling financial market upset
in the second half of 2007 instead of outright recession.
It is also worth noting that the model did not give false recession alarms in previous periods
of financial market turbulence, such as the fall of 1998.
The second chart depicts the four-quarter-ahead recession probabilities.
Four-quarter forecasts help us assess whether the model was able to predict the
duration as well as the occurrence of the 2001 recession.
To avoid false alarms in the late 1990s, the threshold probability level would have
to be almost 40 percent. In this case, the
model did well in calling the second and third quarters of the 2001 recession one year ahead.
In the present case, the model called a recession for 2007Q4, using information available in
2006Q4, when it now looks likely that the first recessionary quarter will be
2008Q1. The model-implied recession probability for 2008Q1, based on data from a year earlier,
exceeds 50 percent, so it appears to be a fairly strong sign of recession.
The third chart uses the model’s capability to forecast at any horizon to generate
out-of-sample recession probabilities, starting with 2008Q1. The first two quarters of
2008 are ones where the model-implied recession probabilities lie above the threshold
level of about 35 percent.
Thus, the model is calling a two-quarter recession for the first half of 2008. Note that the
model-implied recession probabilities dip down near 10 percent in the middle of 2009 and
rise gradually toward the unconditional probability of about 17 percent. The dip below
the unconditional probability is the result of a snap-back phase the economy experiences
after recessions, during which the economy is farther than normal from recessionary conditions.
To get a handle on how unprecedented it would be for the economy to avoid
recession in 2008, it is useful to look at the business cycle index produced by the
Qual VAR model. The fourth chart shows the business cycle index for the entire
sample period with forecasts for 2008 through 2011. The chart shows that
the business cycle index has never dipped as close to zero as it is projected
to do in 2008 without a recession. For this reason, it appears unlikely
that the economy will escape the current slowdown without a recession.