Internet Search Behavior as an Economic Forecasting Tool: The Case of Inflation Expectations

Abstract

This paper proposes a measure of real-time inflation expectations based on metadata, i.e., data about data, constructed from internet search queries performed on the search engine Google. The forecasting performance of the Google Inflation Search Index (GISI) is assessed relative to 37 other indicators of inflation expectations – 36 survey measures and the TIPS spread. For decades, the academic literature has focused on three measures of inflation expectations: the Livingston Survey, Survey of Professional Forecasters, and the Michigan Survey. While useful in developing models of forecasting inflation, these low frequency measures appear anachronistic in the modern era of higher frequency and real-time data. I demonstrate that higher frequency measures tend to outperform lower frequency measures in tests of accuracy, predictive power, and rationality. Furthermore, Granger Causality tests indicate that the GISI metadata indicator anticipates the inflation rate by 12 months, and out-of-sample forecasts show that the GISI has the lowest forecast error of all the inflation expectations indicators tested.

Keywords: Inflation, expectations, metadata, search, surveys, TIPS, Google, forecasting, big data, inflation expectations, out-of-sample forecasts, Federal Reserve

JEL Classification: A10, C12, C13, C22, C42, C51, C52, C53, C81, C82, D12, D83, D84, E31, E44, E58