How Can 27 Million Newspaper Articles Predict GDP? This Zicklin Prof Explains
December 18, 2024
It’s no secret that people make spending decisions based on what they think of the state of the economy. If you have a stable job and you think the economy is in good shape overall, you’re more likely to make a large purchase like a car or a house than, say, someone who fears a layoff or a recession. And because large consumer expenditures are one of many factors that help move the economy, consumer sentiment is an important predictor of economic forecasts. That’s why analysts and policymakers routinely ask businesses and consumers what they think of the state of the economy when making predictions about where the economy is going.
There’s a hitch, though—these business and consumer surveys are released on a monthly basis. Although that’s faster than the GDP report, which is quarterly, it’s still not particularly timely. But what if there were a way to analyze consumer sentiment in almost real time?
That’s what the Zicklin School’s Sebastiano Manzan, PhD, associate professor of economics in the Bert W. Wasserman Department of Economics and Finance, set out to do. In a paper published in the Journal of Applied Econometrics, Dr. Manzan and colleagues from the European Commission Joint Research Centre (a research arm of the European Union) used big data in the form of newspaper articles in major western European countries to see if economic news could predict GDP and other macroeconomic indicators.
“The idea was to look at newspapers to see how they talked about unemployment, inflation, GDP, all this economic news that people read and might use to form expectations about the future,” Manzan explains. “For example, if your company is laying people off, and then you read about layoffs and inflation in [the Italian newspaper] La Repubblica, that might affect your spending decisions.”
To that end, Manzan and his colleagues collected information from—wait for it—27 million articles in 26 major newspapers published between 1995 and 2020 in France, Germany, Italy, Spain, and the United Kingdom. Information from daily news reports, they reasoned, could help analysts take the pulse of economic sentiment more accurately than information released by bureaus and agencies long after the fact.
The first challenge for the team was translating everything into English, which they accomplished using neural machine translation (NMT), a precursor to large language models (LLMs), with Manzan and his multilingual coauthors spot-checking the translations’ accuracy.
Next, they isolated sentences in the articles that discussed economic variables, using keywords from a World Bank lexicon of terms related to economics and monetary policy such as “unemployment,” “inflation,” mortgage rates,” and so on. Sentences were grouped by category and assigned positive or negative sentiments: For example, “slowdown” was negative, but “inflation slowdown,” a double negative, was positive. After analyzing all the information in this enormous dataset, they created six sentiment indicators per country, per day, for the 25 years encompassed in the news reports.
Manzan and his colleagues added the sentiment indicators to a model used to forecast GDP using the most recent available statistics on the economy (unemployment rate, inflation rate, etc.) plus the results of national business surveys issued by the European Commission in each country.
“Once we put all this information into the model, we saw that there was predictive value in the sentiment indicators, specifically regarding inflation, unemployment, and monetary policy,” Manzan explains. “Therefore, we concluded that there is a case for considering sentiment measures calculated from news as an additional tool that economic forecasters should adopt in their quest to produce accurate predictions of economic activity.”
Interestingly, Manzan observes that since his paper was published, both the European Central Bank and the Federal Reserve Bank of San Francisco have incorporated models using national sentiment based on news reports into their forecasts. “Not that I’m taking credit for that!” he laughs. “Many people have been working in this field for a long time.”