Seventeen forecasts and one guess! One of the key issues in the forecast, of course, is the role of the output gap, and I have a few comments to make about that topic. The use of the output gap or the Phillips curve relationship in forecasting inflation was a subject at our two-day meeting in June 2002. At that meeting, Art Rolnick of the Minneapolis Fed presented research by Andrew Atkeson and Lee Ohanian suggesting that output gap measures were no better forecasters of inflation than a simple random walk assumption that inflation next year will equal inflation this year. In particular, Atkeson and Ohanian showed that the Greenbook forecast of inflation, which makes heavy use of the output gap concept, did no better than the random walk model for the period 1984 to 1996.
Given our current reliance on the output gap concept for projecting inflation, I thought it would be worthwhile to revisit this discussion briefly. In particular, I would like to call your attention to recent work by staff members at the Board and at the Federal Reserve Bank of Boston. Work by Board staff member Deb Lindner summarized in a memo last summer, which I’m sure she’d be glad to make available, confirms the results regarding the accuracy of Greenbook forecasts for the period 1984 to 1996 for inflation measured by the GDP deflator. However, she also shows that the Atkeson–Ohanian results regarding the accuracy of Greenbook forecasts are highly fragile on a number of dimensions, notably with respect to the sample period employed and the measure of inflation used. Using the GDP deflator to measure inflation and using real time data only, Lindner shows that when the Atkeson–Ohanian sample period is extended back to 1980, the Greenbook forecasts are 40 percent better in terms of root mean squared errors than the random walk alternative. In the more recent 1997-2002 period—for which we, of course, have the Greenbook data but they are not yet publicly available— Greenbook forecasts of GDP deflator inflation are 35 percent better than the random walk benchmark. Lindner shows that the results are more dramatic still when inflation is measured by the core CPI rather than by the GDP deflator. Even for the Atkeson–Ohanian sample period, 1984 to 1996, Greenbook forecasts for core CPI inflation are 24 percent better than the random walk alternative. For the recent period, 1997-2002, real-time Greenbook forecasts of core CPI inflation are a full 61 percent better than the random walk. These results are consistent with those of several published papers, including a well-known paper by Christina and David Romer that showed that the Greenbook forecasts of inflation have outperformed private-sector forecasts.
The other recent study to which I’d like to call your attention is an article by Boston Fed economists Michelle Barnes and Giovanni Olivei in the most recent New England Economic Review. Barnes and Olivei estimate a piecewise (linear) Phillips curve that allows the effect of the output gap on inflation to vary depending on how far away the economy is from potential. They find little relationship between unemployment and inflation in a region close to full employment but a robust relationship when the unemployment rate is relatively far—which they define as 1.4 percentage points—away from the natural rate. Barnes and Olivei propose economic interpretations of this finding, but I would suggest a measurement-error interpretation. Because our measures of the NAIRU are necessarily quite noisy, when the output gap is small, the measurement noise dominates the signal. Only when the measured output gap is relatively large can we be reasonably sure that an actual output gap exists and, therefore, that inflation will respond.
These results, by the way, are consistent with my own informal investigations of the predictive power of the Phillips curve. For the period from 1960 to the present, I find that a very simple Phillips curve specification does not out-predict the random walk out of sample except at times when unemployment is at least 1 percentage point above its average for the previous six years. In short, output-gap-based forecasts of inflation are probably the most reliable during periods of recession. And indeed, inflation behaved as the output gap theory would suggest in 2002 and 2003. Looking forward to 2004, the unemployment rate is now less than 1 percentage point away from the estimated value of the NAIRU and from its recent average. While a continuing output gap may induce some further disinflation, from a statistical point of view the random walk model may now be just as good. That is, following President Stern, a good guess is that inflation in 2004 will be the same as in 2003.
With your permission, I’d like to add one thought on the “considerable period” language that we’ll be discussing later. A good rule of thumb is to try to look as if you know what you’re doing even if you’re not entirely sure. We properly emphasized that “considerable period” refers to economic time, not calendar time, and we made our commitment explicitly conditional on low inflation and resource slack. We can debate whether or not the intermeeting data, including the December jobs report and very low inflation numbers, suggest improvement on those two dimensions. However, the bond markets clearly believe that they do not, as yields have fallen significantly and the expected date of Fed tightening has been pushed further into the future. Hence, as our conditionality is not perceived to have been satisfied, we have no fig leaf for dropping the “considerable period” language today. I would rather wait until March and the presumption that we will see at least one good payroll number by then. In short, I’m looking now more at long-term credibility issues rather than short-term flexibility and tactical issues. Of course, if we don’t see a strong payroll number by March, then we might be glad that we didn’t drop the language. Thank you.