Predicting Unemployment Claims Using Regional and Exogenous Signals: A Sparse Modeling Approach

Abstract

In this paper we apply a time series based Vector Auto Regressive (VAR) approach to the problem of predicting unemployment insurance claims in different census regions of the United States. Unemployment insurance claims data, reported weekly, are a leading indicator of the US unemployment rate. Gathering weekly unemployment claims and aggregating by region, we model correlation between the different census regions. Additionally, we explore the use of external variables such as Bing search query volumes and URL site clicks related to unemployment claims. To prevent any spurious predictors from appearing in the model we use sparse model based regularization. Preliminary results indicate that our approach is promising and in ongoing work we are extending the approach to a larger set of predictors and a longer data range.