Individuals who find themselves jobless frequently share their experiences prior to submitting any official documents. They express their frustrations, seek referrals from their networks, or simply declare their unemployment status. These disclosures, dispersed among millions of social media users, have been shown to forecast formal unemployment claims up to fourteen days ahead of the government’s data release.
Academics from New York University, the World Bank, and the University of Oxford developed an AI model that examines Twitter for announcements of job loss. They trained this model on 31.5 million U.S. users’ posts from 2020 to 2022. This system, referred to as JoblessBERT, identifies informal expressions that conventional keyword searches might overlook—such as “needa job” or “neeeeeed a job!” It detected almost threefold more unemployment-related posts than previous techniques, while maintaining accuracy.
The model does not evaluate emotions or perspectives regarding the economy. It highlights explicit claims: I lost my job. I’m unemployed. I need work.
Addressing the Twitter-Isn’t-Everyone Issue
Twitter users tend to be younger than the broader population. Not every individual who experiences job loss makes a post about it. The researchers estimated each user’s age, gender, and location, and subsequently adjusted their counts to align with U.S. Census demographics. They refer to this as post-stratification. With these adjustments, they were able to predict unemployment insurance claims at national, state, and city levels.
The AI utilized active learning, enhancing its capabilities by concentrating on uncertain instances where a post may or may not indicate job loss. Gradually, it encompassed a more diverse array of users across different demographics and regions.
In March 2020, unemployment claims surged from 278,000 to nearly 6 million within a fortnight as COVID-19 led to widespread economic shutdowns. Government systems struggled to handle the sudden influx. Two days before the official reporting week concluded, JoblessBERT forecasted 2.66 million claims. The actual figure was 2.9 million. Industry analysts employing traditional forecasting methods significantly underestimated this surge.
“We demonstrate that our approach consistently exceeds the industry’s prediction average and can enhance the forecasts of U.S. unemployment insurance claims, up to two weeks ahead,” states Samuel P. Fraiberger of the World Bank Development Impact Group.
Throughout the entire study duration, forecasts based on social media signals reduced prediction errors by more than 50 percent in comparison to industry consensus.
Timely Alerts Amidst Economic Turmoil
The benefit is particularly evident during swift changes. When circumstances shift quicker than data gathering cycles permit, an advance notice of two weeks is crucial. Policymakers can allocate resources, modify programs, or prepare communications while conventional statistics are still being compiled.
In ten cities where official data updates are scarce or absent, the model still generated dependable estimates. The digital signals remain effective even without local government benchmarks for comparison.
The authors do not propose replacing official labor statistics. Conventional surveys continue to be extensive and methodologically sound. However, they offer a real-time complement. What individuals are experiencing presently, rather than what they reported in last month’s survey.
The limitation lies in data access. Social media platforms have made it more challenging for researchers to access public posts in recent years, complicating the sustained effort of this type of work. If researchers can utilize anonymized data responsibly while safeguarding privacy, these signals can become instruments for public interest analysis. Currently, the next series of job losses may surface in social feeds long before making its way to official spreadsheets. Whether anyone can still access it remains an open question.
PNAS Nexus: 10.1093/pnasnexus/pgaf309
There’s no paywall here
If our reporting has either informed or inspired you, we kindly ask you to consider contributing. Every donation, regardless of its size, empowers us to continue providing accurate, engaging, and trustworthy science and medical news. Independent journalism demands time, effort, and resources—your support ensures we can keep revealing the stories that matter most to you.
Join us in making knowledge accessible and impactful. Thank you for your support!