Generating Up-to-date Starting Values for Detailed Forecasting Models

Abstract

In economic forecasting, it is important that the forecasts be based on data that is both reliable and up-to-date. The most reliable data typically come from conducting a census. These censuses produce estimates with a long lag between the reference year and the date of publication. However, we also have other sources of economic data that are less reliable but published more frequently. These higher frequency data should be a source of useful information for analyzing economic activity in the current, incomplete year.

The objective of this study is to use high frequency (monthly and quarterly) data to generate forecasts of the annual data from reliable sources used in an inter-industry forecasting model. The results will be used as starting values to improve the model's short-term forecast performance.

The distinguishing feature of this dissertation is that it studies the economic data at the sectoral level as opposed to other studies that only try to generate aggregate data. The aggregate data will be a by-product of these detailed estimates. Thus, we can forecast the trends of the aggregates and observe sectors that contribute to these trends.

In this dissertation, I study data on four main aspectts of the U.S. economy: 1) Personal consumption expenditures, 2) Investment in equipment and software, 3) Investment in structures, and 4) Gross output.

By historical simulations, I find that the performance of the forecasts depends heavily on the accuracy of the exogenous variables used in each forecast. The estimated detailed values are consistent with the macroeconomic data, used as regressors in the processes. Thus, generally, the results will be reliable as long as we have a good forecast of macroeconomic variables.

The performance of the first-period forecast also depends on where in the calendar year the last published data is. The closer to the end of the year, the better is the accuracy of the forecast.