During the U.S. economic recovery from the COVID-19 induced economic recession, many observers noted that metrics of consumer sentiment showed that consumers held very negative feelings about the economy despite strong metrics for unemployment and real GDP growth. This negative consumer sentiment persisted even when inflation moderated and economists found that sentiment was negative compared to prior periods where the economy was performing similarly.
Several mechanisms for this disconnect were proposed. One mechanism, proposed by Ben Harris and Aaron Sojourner of Brookings, is the negative tone of media coverage of the economy. Harris and Sojourner found that when controlling for economic fundamentals, media's coverage of the economy became more negative in 2017 and remained negative through 2023.
This page replicates the Harris and Sojourner analysis and extends it to more recent quarters. This page will update quarterly once GDP growth data is available from the Bureau of Economic Analysis. There are two charts on this page. The first, below, shows the entire data series. After looking at different hold-out periods to test their model out-of-sample, Harris and Sojourner settled on 2017 as a cut point for the model. The data suggests that from 2017 onwards, there was a shift in how journalists reported on the economy, with coverage starting to skew more negatively. I adopt the same model they use. Data before 2017 is used to train an ARIMA model to predict media tone based on year-over-year changes in the stock market, inflation, GDP growth, and the current unemployment level. Observed media tone and predictions from the model are shown below.
The second chart, below, looks more closely at the out-of-sample data from 2017 onwards. The last quarter included in Harris and Sojourner's analysis was Q4 of 2023. I update their analysis and show that media's tone on the economy has remained more negative than what the 1988-2017 training set predicts. This graph shows only the difference between actual tone and predicted tone. Yellow bars indicate quarters where reporting on the economy was more positive than the model predicted, and red bars indicate quarters where reporting was more negative.
You can find my code for this project, which replicates Harris and Sojourner's analysis in Python, on my github. Harris and Sojourner's original Stata analysis is available on Sojourner's github.