Economic Measurement Lost in a Random Forest? A Case Study of Employment Data

Big data and machine learning (ML) offer transformative potential for economic measurement. This study evaluates the use of alternative employment data from a payroll processor to improve on timely measures of regional employment estimates, comparing ML methods--Lasso regression and Random Forest (RF)--to linear models. RF models show substantial improvements in cross-validation but struggle with extrapolation, particularly during the pandemic. At the county level, greater data variation aids prediction, though sampling errors complicate performance. These findings highlight ML's promise in improving economic statistics while emphasizing the need for careful model selection, robust evaluation metrics, and consideration of data-specific challenges.