Learning the Structure of Price Stickiness in Scanner Data

Abstract

A substantial portion of the macroeconomics literature suggests that demand shocks have noteworthy effects on real output. One prevailing theory for rationalizing the large effects of demand shocks on output is that nominal prices are not perfectly flexible. Our work aims to extend the empirical literature studying price stickiness by using models and techniques from Bayesian statistics and machine learning. We analyze a high-frequency scanner dataset and focus on algorithmically identifying price setting periods for any given product and characterizing a set of regular prices within price setting periods, given only the values of the price time series. Our work departs from the existing literature in three important ways. For each product, we flexibly identify price setting periods without making any assumptions on the length of the periods or the number of periods. Additionally, we do not simply study one regular, or reference, price. Instead, we identify several "modal" prices from the data that correspond to the peaks of the price distributions for the identified price setting periods. Finally, we develop highly clusterable and interpretable metrics for price stickiness, which we demonstrate with an easily separable clustering of the products into two clean groups of highly sticky and non-sticky products.