Nowcasting with Numerous Candidate Predictors

Abstract

The goal of nowcasting, or “predicting the present,” is to estimate up-to-date values for a time series whose actual observations are available only with a delay. Methods for this task leverage observations of correlated time series to estimate values of the target series. This paper introduces a nowcasting technique called FDR (false discovery reduction) that combines tractable variable selection with a time series model trained using a Kalman filter. The FDR method guarantees that all variables selected have statistically significant predictive power. We apply the method to sales figures provided by the United States census bureau, and to a consumer sentiment index. As side data, the experiments use time series from Google Trends of the volumes of search queries. In total, there are 39,059 potential correlated time series. We compare results from the FDR method to those from several baseline methods. The new method outperforms the baselines and achieves comparable performance to a state-of-the-art nowcasting technique on the consumer sentiment time series, while allowing variable selection from over 250 times as many side data series.