News Sentiment in the Time of COVID-19

The COVID-19 pandemic is causing severe disruptions to daily life and economic
activity in the United States and around the world. These ruptures were
immediately evident in financial markets, with equity prices declining sharply
and market volatility spiking. It is also readily apparent that consumer
spending in sectors like leisure and hospitality is falling dramatically due
to shelter-in-place and other social distancing measures being imposed across
the country.

Assessing the timing and magnitude of the economic fallout in this rapidly
evolving situation has been hampered by the low frequency and lagged
availability of most macroeconomic data. In particular, so-called hard data
such as payroll employment, personal income, consumer spending, and business
investment are published with lags of weeks or months. Analysts and
policymakers are particularly interested in how consumer and business
sentiment is holding up right now given the well-documented links between
sentiment and economic activity (see, for example, Carroll, Fuhrer, and Wilcox
1994, Benhabib and Spiegel 2020, and Shapiro and Wilson 2017). Available
survey-based sentiment indexes are either low frequency, which limits their
usefulness in times of sudden change, or have a short history, which prevents
comparisons with past episodes.

In this Letter, we discuss the newly developed Daily News
Sentiment Index that provides real-time data from 1980 to today. The index was
developed and analyzed in Shapiro, Sudhof, and Wilson (2020). This daily index
is highly correlated historically with the monthly survey-based
University of Michigan Index of Consumer Sentiment and the Conference Board’s
Consumer Confidence Index.

Our Daily News Sentiment Index began falling sharply in January of this year,
coinciding with increasing news coverage of the coronavirus disease 2019
(COVID-19). This change appeared two months earlier than in the survey-based
sentiment measures. The decline in March was especially steep, consistent with
the large drop in consumer sentiment indexes in March.

Measuring news sentiment

Sentiment analysis is a rapidly developing field of natural language
processing and is now widely used in an array of business applications, such
as social media, algorithmic trading, customer experience, and human resource
management. The process allows one to directly quantify the emotional content
from any set of text. There are two general approaches. The first, known as
the lexical approach, relies on a predefined list of words associated with an
emotion, referred to as lexicons. For example, sentiment lexicons typically
classify words into three categories: negative, neutral, or positive. The
second, more nascent approach relies on machine-learning (ML) techniques to
predict the sentiment of a given set of text. ML techniques can, in principle,
learn sentiment weights for words and even entire phrases, then use those
weights to measure the sentiment of the given textual passage. The drawback of
the ML approach is that it requires large training data sets labeled for the
terms that are specific to the domain of interest—for example, business texts
or social media posts—which are time-consuming and expensive to construct.

The study by Shapiro, Sudhof, and Wilson (2020, hereafter SSW), constructs
sentiment scores for economics-related news articles using a lexical approach.
It uses a historical archive of news articles from 16 major U.S. newspapers
compiled by the news aggregator service LexisNexis. The newspapers cover all
major regions of the country, including some with extensive national coverage
such as the New York Times and the Washington Post.
SSW selected articles with at least 200 words where LexisNexis identified the
article’s topic as “economics” and the country subject as “United States.”
Combining publicly available lexicons with a news-specific lexicon constructed
by the authors, the study develops a sentiment-scoring model tailored
specifically for newspaper articles.

To assess the model’s accuracy, the authors compared sentiment scores from
this hybrid lexical model with human-provided sentiment scores for a random
sample of 800 news articles. These latter scores were generated by a team of
research assistants at the Federal Reserve Bank of San Francisco, who were
asked to rate articles on a scale of 1 to 5, from very negative to very
positive. The SSW hybrid lexical model was strongly correlated with the human
scores, performing better than any single lexical model and similar to or
better than models constructed using existing machine-learning techniques.

Comparing news sentiment with survey-based consumer sentiment

SSW next aggregated the individual article scores into daily and monthly
time-series measures of news sentiment, relying on a statistical adjustment
that accounts for changes over time in the composition of the sample across
newspapers. In this Economic Letter, we extend and augment the
SSW daily sentiment measure in two ways. First, we update the set of news
articles from LexisNexis to the present, since the news archive used in SSW
was only through mid-2015. Second, because the day-to-day fluctuations in the
sentiment measure tend to be noisy, we construct a smoothed daily index as a
trailing weighted average of the raw data, with weights that decline
geometrically with the length of time since the article’s publication. This
weighted average is analogous to how capital stocks are generally measured
from past vintages of investment: older investments contribute less according
to an assumed depreciation rate. We assume a depreciation rate of 5%; that is,
with each passing day, articles become 5% less relevant for today’s sentiment.
Our results in this Letter are not particularly sensitive to the exact
depreciation rate used.

Figure 1 shows the resulting Daily News Sentiment Index over time (blue line).
Updates to this index are provided regularly on the San Francisco Fed’s
new data page. Figure 1 also includes the University of Michigan’s Index of Consumer
Sentiment, which is derived from a survey and is available at a monthly
frequency (green line). For both indexes, higher values indicate more positive
sentiment. Though the units of the two indexes are not directly comparable, it
is interesting to consider how each index has moved in relation to the
business cycle and around the time of major events. It is also interesting how
each index has moved recently, compared with responses to past events.

The news sentiment index correlates strongly with the survey-based consumer
sentiment measure, and both are strongly procyclical, dipping during
recessions and rising during economic expansions. We found similar results
using the Conference Board’s Consumer Confidence Index (not shown). The news
sentiment index also tends to move with key historical events that have
affected economic outcomes and financial markets, such as the start of the
first Gulf War in August 1990; the Russian financial crisis in August 1998;
the terrorist attacks of September 11, 2001; the Lehman Brothers bankruptcy in
September 2008; and the October 2013 federal government shutdown. So far in
the recent daily results, the news sentiment index has not yet fallen to the
low points of the past three recessions, though it may well fall further in
the days and weeks ahead.

Sentiment in the time of COVID-19

Figure 2 zooms in on the last 18 months of data from Figure 1 to focus on the
most recent movements in sentiment. The Michigan Consumer Sentiment Index
remained elevated through February. However, March saw the fourth largest
one-month decline in the index since 1980. The Daily News Sentiment Index, on
the other hand, shows a sharp drop in sentiment beginning in early January and
steepening further in the first two weeks of March.

How much of the drop in news sentiment in recent weeks can be attributed to
the COVID-19 outbreak? To address this question, Figure 3 shows our index
(dark blue line) along with a measure of COVID-19 news coverage (light blue
line). For the latter, we calculate the fraction of economics-related news
articles that contain the terms “coronavirus” or “COVID-19.” The series in
Figure 3 is a trailing-average of this fraction, with weights that decline
geometrically for older articles, analogous to how we constructed the news
sentiment index. Articles mentioning the coronavirus and COVID-19 began around
January 20 and then rapidly increased. By late March, the percentage of
economics-related news articles mentioning the virus reached an astounding
95%. The figure clearly shows that the decline in sentiment through mid-March
coincided with the increased coverage of COVID-19. More recently, the news
sentiment index has flattened, coinciding with the rise in news coverage
related to fiscal stimulus legislation (green line). There was also a
temporary decline in sentiment in the first half of January, which was due to
the flare-up of U.S.-Iran hostilities and related disruptions in the oil
market.

Conclusion

The new Daily News Sentiment Index introduced in this Letter can
be especially useful in times of sudden economic change as we are experiencing
now. This index is constructed at a daily frequency from 1980 to today. By
contrast, existing survey-based sentiment indexes have either low frequency or
a short history. The Daily News Sentiment Index is highly correlated with
survey-based measures of consumer sentiment, but it shows an earlier and more
pronounced drop in recent weeks. This rapid decline in news sentiment has
coincided with the increasing news coverage of COVID-19.

Studies have documented a strong link between sentiment and subsequent
economic activity, especially business investment and consumer spending. For
example, Benhabib and Spiegel (2019), using state-level data, find that a
one-standard-deviation drop in consumer sentiment as measured by the Michigan
survey would be expected to result in a 2.3% drop in personal consumption
expenditures. The drop over the past two months in the Michigan Consumer
Sentiment Index has, in fact, been very close to one standard deviation. It is
therefore worth noting, for comparison, that our Daily News Sentiment Index
has fallen over the past two months by a little over three standard
deviations, which may portend a steep decline in consumer spending. It is
important to note that the current episode is particularly unique. Consumer
spending at the moment has been depressed for reasons beyond low sentiment. In
particular, government-imposed social distancing measures are directly
inhibiting spending on many types of discretionary goods and services, such as
those associated with leisure and hospitality.

In the weeks and months ahead, it will be important to monitor news and
consumer sentiment to see how much further sentiment will fall and when it
will start to turn around.

Shelby R. Buckman is research associate in the Economic Research Department of
the Federal Reserve Bank of San Francisco.

Adam Hale Shapiro is
research advisor in the Economic Research Department of the Federal Reserve
Bank of San Francisco.

Moritz Sudhof is cofounder and chief executive officer at Motive Software.

Daniel J. Wilson is
vice president in the Economic Research Department of the Federal Reserve Bank
of San Francisco.

References

Benhabib, Jess, and Mark M. Spiegel. 2019. “Sentiments and Economic Activity:
Evidence from U.S. States.” The Economic Journal 129(618,
February), pp. 715–733.

Carroll, Christopher D., Jeffrey C. Fuhrer, and David W. Wilcox. 1994. “Does
Consumer Sentiment Forecast Household Spending? If So, Why?”
American Economic Review 84(5), pp. 1,397–1,408.

Shapiro, Adam H., Moritz Sudhof, and Daniel J. Wilson. 2020.
“Measuring News Sentiment.”
FRBSF Working Paper 2017-01.

Shapiro, Adam, and Daniel J. Wilson. 2020.
“What’s in the News? A New Economic Indicator.”
FRBSF Economic Letter 2017-10 (April 10).