Title: Now- and backcasting initial claims with high-dimensional daily internet search-volume data
Authors: Daniel Borup - Aarhus University (Denmark) [presenting]
David Rapach - Washington University in Saint Louis and Saint Louis University (United States)
Erik Christian Montes Schutte - Aarhus University (Denmark)
Abstract: A sequence of now- and backcasts of weekly unemployment insurance initial claims (UI) is generated based on a rich trove of daily Google Trends (GT) search-volume data for terms related to unemployment. To harness the information in a high-dimensional set of daily GT terms, we estimate predictive models using machine-learning techniques in a mixed-frequency framework. The sequence of now- and backcasts are made ten days to one day before the release of the UI figure on Thursday of each week. In a simulated out-of-sample exercise, now- and backcasts of weekly UI that incorporate the information in the daily GT terms substantially outperform those based on an autoregressive benchmark model, especially since the advent of the COVID-19 crisis. The improvements in predictive accuracy relative to the autoregressive benchmark generally increase as the now- and backcasts include additional daily GT data, with reductions in root mean squared error of up to approximately 50\%. Variable-importance measures reveal that the GT terms become more relevant for predicting UI during the crisis. At the same time, partial-dependence plots indicate that linear specifications are largely adequate for capturing the predictive information in the GT terms. We are in the process of creating a website that will provide updated, real-time now- and backcasts of UI daily.