Forecasting inflation using sentiment

Using algorithmically scored sentiment of almost 730.000 news articles between Q1 2003 and Q4 2021, we construct an index and analyze its predictive power for US inflation for up to eight quarters. In a pseudo out-of-sample setting, we show that sentiment is able to forecast inflation more accurately than a naïve random walk with root mean squared errors that are around 30 percent lower depending on the forecasting horizon. Against other often used benchmarks, forecasting models using macroeconomic variables and Michigan surveys, forecasting accuracy of our sentiment index tends to outperform for shorter forecasting horizons. •We construct an index based on news sentiment to forecast US inflation•Over 700’000 news articles are algorithmically examined for topics•Sentiment is able to forecast inflation more accurately than a naïve random walk•Sentiment outperforms other survey-based sentiment indicators over shorter horizons