Forecasting with Real-Time Macroeconomic Data

Dean D. Croushore, University of Richmond


Forecasts are only as good as the data behind them. But macroeconomic data are revised, often significantly, as time passes and new source data become available and conceptual changes are made. How is forecasting influenced by the fact that data are revised?

To answer this question, we begin with the example of the index of leading economic indicators to illustrate the real-time data issues. Then we look at the data that have been developed for U.S. data revisions, called the “Real-Time Data Set for Macroeconomists” and show their basic features, illustrating the magnitude of the revisions and thus motivating their potential influence on forecasts and on forecasting models. The data set consists of a set of data vintages, where a data vintage refers to a date at which someone observes a time series of data; so the data vintage September 1974 refers to all the macroeconomic time series available to someone in September 1974.

Next, we examine experiments using that data set by Stark and Croushore (2002), Journal of Macroeconomics 24, 507–531, to illustrate how the data revisions could have affected reasonable univariate forecasts. In doing so, we tackle the issues of what variables are used as “actuals” in evaluating forecasts and we examine the techniques of repeated observation forecasting, illustrate the differences in U.S. data of forecasting with real-time data as opposed to latest-available data, and examine the sensitivity to data revisions of model selection governed by various information criteria.

Third, we look at the economic literature on the extent to which data revisions affect forecasts, including discussions of how forecasts differ when using first-available compared with latest-available data, whether these effects are bigger or smaller depending on whether a variable is being forecast in levels or growth rates, how much influence data revisions have on model selection and specification, and evidence on the predictive content of variables when subject to revision.

Given that data are subject to revision and that data revisions influence forecasts, what should forecasters do? Optimally, forecasters should account for data revisions in developing their forecasting models. We examine various techniques for doing so, including state-space methods.

The focus throughout this chapter is on papers mainly concerned with model development – trying to build a better forecasting model, especially by comparing forecasts from a new model to other models or to forecasts made in real time by private-sector or government forecasters.