An interesting context of the publications: the outbursts, the requests for recommendations – even before they were unemployed.
The analysis was carried out in the USA but can be applied to many other countries: social media posts predict an increase in unemployment (two weeks earlier).
People vent publicly, leave public requests for help, for recommendations – even when they are not yet officially unemployed. Or, simply, they announce that they are out of a job.
These publications spread through social networks and researchers argue that they can function as a early warning system for unemployment spikes — up to two weeks before the official Government figures are released.
The applied an artificial intelligence model that analyzes the X looking for straightforward publications about unemployment.
The system, called JoblessBERT, was trained with publications from 31.5 million users in the United States between 2020 and 2022; was designed to detect informal formulations and a lot variables that simple keyword searches tend to fail — including misspellings, slang, and emphatic expressions.
The explains that rather than measuring mood or broader economic sentiment, the model focuses narrowly on explicit statements that indicate job loss or need for work.
In testing, researchers report that JoblessBERT identified almost three times more relevant publications than previous approaches while maintaining accuracy.
Because X users are not representative of the general population — they tend to be younger and exclude many people who never post about employment — the team used demographic and geographic inference to estimate each user’s age, gender, and location.
It then aligned with US Census proportions through a statistical correction known as post-stratification. With this step, they say, the system can predict unemployment benefit claims not only nationally, but also by states and cities.
The model also used active learning (active learning), a technique that improves performance by focusing training on ambiguous cases, in which it is not clear whether or not a publication indicates job loss.
Over time, this helped capture a broader sample of language and users, spanning distinct regions and demographic groups.
The approach was tested in the first weeks of the “shock”. In March 2020, initial claims for unemployment benefits in the US soared from 278,000 to nearly 6 million in two weeks as lockdowns spread across the economy. Two days before the official reporting week ended, JoblessBERT projected 2.66 million applications; the final number was 2.9 million — a much closer estimate than many traditional industry forecasts, which significantly underestimated the increase.
According to researchers, combining social media signal with standard prediction methods reduced forecast errors by more than 50% against industry consensus throughout the study period. Gains have been most visible in phases of rapid change, when conventional data collection lags behind.
The authors emphasize that the model is not intended to replace official statistics of the labor market, which continue to be more comprehensive and methodologically robust. Instead, they present it as a real-time supplement — capable of capturing what people report immediately, and not what only emerges later in surveys or administrative records.
A big obstacle, however, it is the access to the data. In recent years, social platforms have tightened restrictions on the collection of public posts by researchers, raising questions about whether these early warning systems can be sustained at the necessary scale, even with privacy safeguards and anonymized analysis.
