Do Leading Indicators Help Predict GDP Growth Rates?
The second estimate for real GDP growth in the fourth quarter of 2011 came in at 3 percent, an increase from the previously estimated 2.8 percent. The acceleration from its 1.7 percent growth rate in the third quarter was mainly due to increases in private domestic inventories, durable goods consumption, and residential investment. This latest estimate leaves real GDP growth for 2011 (in annual levels) at 1.7 percent, far below the 2010 growth rate of 3 percent, and further emphasizes the uneven nature of the current recovery.
This high degree of uncertainty surrounding the recovery has economists grasping at straws when it comes to explaining short-term growth. Each new data release seems to be of the utmost importance, and each new speech by an economic principal is carefully parsed as if it were an oracle’s. The Conference Board’s Leading Economic Index (LEI), an aggregate of 10 leading indicator data series, is closely watched for possible turning point patterns in output. Typically, the index tends to start declining before recessions and to come back up before the recession is over, signaling a recovery.
Apart from their ability to predict turning points, do leading indicators carry any information besides what is already embedded in past real GDP that can be useful in forecasting GDP growth rates in the next quarter? To answer this question, I looked at the LEI, as well as at some of the individual series that it uses.
In addition to the LEI, I looked at a total of eight series, starting in 1985:
- Average weekly hours worked in manufacturing
- The spread between the 10-year treasury note rate and the federal funds rate
- The S&P 500 index
- consumer expectations from the Michigan University survey
- Manufacturers’ new orders of nondefense capital goods
- The new manufacturing orders index from the Institute for Supply Management (ISM)
- Weekly new unemployment insurance benefit claims
- New private housing units building permits.
I investigated whether the quarterly behavior of these indicators somehow precedes that of output as given by real GDP. Since some of these variables grow over time, like real GDP or the S&P 500, I took quarterly differences so that I am in fact dealing with changes. (Note that the LEI was recently updated. It now excludes aircraft orders from manufacturers’ new orders of nondefense capital goods. Added to the index are the Leading Credit Index, a proprietary indicator, and manufacturers’ new orders of consumer goods and materials.)
Because I am assessing whether past values of each of the leading indicators provide explanatory power beyond that of past GDP changes, I first removed the variation in current GDP changes that past GDP changes can account for. I did that by regressing this quarter’s GDP growth on past quarters’ GDP growth. Including more than 2 quarters of past data is not worthwhile. I then tested whether the remaining change in current GDP could be explained with changes in the past behavior of the leading indicators. One big caveat is in order here: I am using revised data, the latest vintage that I had available. Practitioners will be working with real-time data, which will, with some exceptions, get revised.
The only variables that seem to have any individual predictive power beyond that already embedded in lagged GDP changes are the S&P500, the ISM’s new manufacturing orders, and private housing building permits. Given that these variables account for less than 25 percent of the composite index, I was then surprised to find, in another experiment I ran, that the LEI does a better job at predicting next quarter’s real GDP changes than lagged values of GDP changes themselves. The figure shows its performance since 1985.
What explains this remarkable accuracy? The Conference Board just revised the weights attributed to each of the component series used in the LEI, and they were computed using 1984 to 2010 data, practically the same sample used in the figure. So the figure is showing in-sample results, in statistical jargon. As Diebold and Rudebusch (1991)* pointed out, if one uses out-of-sample, real-time series, the performance of the index should deteriorate significantly. (Note I do not produce the out-of-sample prediction, since I don’t have data with earlier LEI weightings.) So while LEI may look like a very good predictor, one should proceed with caution.
*“Forecasting Output with the Composite Leading Index: A Real-time Analysis,” Francis X. Diebold and Glenn D. Rudebusch, 1991. Journal of the American Statistical Association, vol. 86, no. 415 (Applications and Case Studies).