COVID-19: Forecasting with Slow and Fast Data
As Yogi Berra is said to have put it, “It’s tough to make predictions, especially about the future.” In the context of economic forecasting, this quote is true even in the best of times, but is even more so in the chaos that is the current economic environment.
As an example, consider the forecast of current quarter real GDP growth implied by the St. Louis Fed Economic News Index (ENI). It currently predicts that the economy will grow 2.26% on an annualized basis in the first quarter.
That is nonsense. As the coronavirus pandemic worsens, many U.S. cities and states are implementing mandatory shelter-in-place orders, which significantly reduce economic activity. Even in areas without such orders, businesses are temporarily shutting down of their own accord. The economy almost certainly will not grow at that rate this quarter and may even contract.
If the ENI is so bad, why do we post it in FRED? The answer is that it isn’t. Historically, the index has actually performed quite well. The problem is that it is spectacularly misinformed right now. Much of the data being fed into it is based on how the economy was performing in January and February. In February, for example, the unemployment rate was at a near-record low of 3.5%. With inputs like that, a prediction of 2.26% doesn’t seem unreasonable.
Lags in Data
The problem isn’t the model per se; it’s that it was constructed to operate using slow data, or data gathered at a relatively low frequency (monthly, quarterly, etc.). In fact, all of the inputs for the ENI are monthly data with a lag that ranges between zero and five weeks.
This means that, as of March 31, the most recent data is based on measurements taken in February. In normal times, that is perfectly adequate for forecasting, since U.S. economic growth rarely exhibits wild, unpredictable movements like it is experiencing now. March data will get included in the ENI but not until it is released in April. Put differently, it may be a few weeks before the ENI starts providing us with forecasts of first-quarter GDP growth that are anywhere near reasonable.
Slow Data Vs. Fast Data
Obviously, slow data is not helping us right now. We need to start looking at fast data: data arriving at a daily or weekly frequency. The problem with fast data is that while it can be informative, it can also be noisy.
A case in point is the stock market. On a daily basis, it can exhibit wild gains and losses that do inform us about the state of the economy, but the signal is buried under the noisy animal spirits of individual investors. When working with higher frequency data, the trick is to find data that are informative without being too noisy.
Fortunately, necessity is the mother of invention. Last week, Daniel Lewis of the New York Fed, Karel Mertens of the Dallas Fed and Jim Stock of Harvard University began posting semiregular forecasts of current quarter GDP growth using a Weekly Economic Index (WEI) that uses high frequency weekly data to form predictions.
Among the 10 weekly series they use as inputs, initial and continued unemployment insurance claims are the only two you have likely heard of. The other eight are a little outside-the-box:
- Federal taxes withheld
- Railroad traffic
- Redbook same-store sales
- Rasmussen Consumer Confidence
- The American Staffing Association Staffing Index
- Steel production
- Wholesale sales of gasoline, diesel and jet fuel
- Weekly average US electricity load
As of March 31, the WEI indicated that GDP would decline by 3.04% at an annualized rate in the first quarter, a much more sensible forecast than that which is currently indicated by the ENI.
- St. Louis Fed’s COVID-19 resource page
- Liberty Street Economics: Monitoring Real Activity in Real Time: The Weekly Economic Index
- On the Economy: COVID-19: What Do FREDcast Users Think about Economic Growth?
- On the Economy: The St. Louis Fed’s Financial Stress Index, Version 2.0