Nowcasting U.S. Inflation: The Role of Online Retail Prices

Abstract: We examine whether high-frequency online retail price data contributes to more accurate nowcasts of the U.S. inflation rate, as given by the monthly change in the Consumer Price Index, when other commonly considered variables for predicting inflation already have been taken into account. This is done by employing three separate models which are all able to handle mixed-frequency data, the Mixed-Data Sampling (MIDAS) regression, the Dynamic Factor Model (DFM) and the Bayesian Vector Autoregression (BVAR), as well as multiple nowcast combination schemes. Following the procedure proposed by Modugno (2013), we are additionally able to disentangle the model-based news from each data release within the DFM framework and thus evaluate how different groups of variables contribute to nowcasting accuracy as data accumulates. The empirical results show that, in most cases, equal predictive accuracy between specifications including and excluding online retail prices cannot be rejected, indicating that this data carries little or no valuable information in excess of other variables, which is robust to varying model specifications and out-of-sample periods. Instead, it is found that commodity prices are particularly good predictors of the U.S. inflation rate and that once these are included, the remaining variables make limited improvements to accuracy.