We find several features of the empirical results surprising and intriguing. Few theoretical macroeconomic models that suggest a linear factor structure for the overall macroeconomy, yet we find that six factors account for almost one-half of the variance of the 170 time series in our balanced panel and twelve factors account for almost two-thirds of this variance. Even if a factor structure describes the joint behavior of these series, there is no reason why a forecast based on static factor estimates should outperform forecasts based on leading indicators or other specialized models that have been fine tuned through years of experience.

Yet, forecasts based on just the first six factors perform well for both CPI inflation and industrial production growth, series that measure quite different economic concepts (nominal prices, real output) and have quite different univariate time series properties. Thus, these results raise numerous issues for future empirical and theoretical research.
One such issue, only touched on in section 5, is the interpretation of the estimated factors.

A feature of traditional diffusion indexes is that they were constructed to have a ready interpretation, such as a measure of how widespread employment growth is across sectors in the economy. Like traditional diffusion indexes, our estimated dynamic factors are averages of many different economic series, but they are identified only up to a nonsingular kxk transformation. Thus the estimated factors will not in general have the natural interpretation that is a feature of traditional diffusion indexes. This raises the question of how to transform the factors into interpretable diffusion indexes. For work on this topic, see Quah and Sargent (1993) and Forni and Reichlin (1996, 1997, 1998).

Several methodological issues remain. One is to explore estimation methods that might be more efficient in the presence of heteroskedastic and serially correlated uniquenesses. Another is to develop a distribution theory for the estimated factors that goes beyond the consistency results shown here and provides measures of the sampling uncertainty of the estimated factors.