645 – Recent Developments in Benchmarking and Reconciliation of Large Systems of Time Series Data: Theory and Practice
Diagnostics for Benchmarking Economic Time Series at the U.S. Census Bureau
Irene Brown
U.S. Census Bureau
A common problem faced by government agencies that collect and publish time series data is to maintain a consistent set of time series. Benchmarking refers to methods used to adjust more frequent time series (e.g. monthly) to match the less frequent series (e.g. annual). Economic Programs at the U.S. Census Bureau recently began using new software to implement the Causey-Trager and Fagan benchmarking methods. Both methods use the iterative, nonlinear constrained optimization technique of steepest feasible descent to obtain a revised series. During parallel testing of thousands of revised series, the need for easy and effective diagnostics became apparent. The two software programs converged to the same solution, except for a small number of series. Further investigation showed that steepest feasible descent had sensitivity to numerical errors for these series. This paper investigates diagnostics, to help identify the types of series that need further review, through an empirical study. We examine graphical analysis suggested by Statistics Canada along with a few alternatives.