The created model, named JoblessBERT, is trained to analyze messages on Twitter (X), where users share information about job loss. Unlike traditional methods that search for specific phrases, the neural network recognizes conversational language, slang, and typos, allowing it to identify 13 times more instances of job loss with high accuracy.
To compensate for the discrepancy between the Twitter audience and the overall population structure of the U.S., the researchers identified the age, gender, and location of users based on profile pictures and adjusted the data based on census results.
The model was tested on data from 2020-2022, including the COVID-19 pandemic period. For example, in March 2020, the number of unemployment benefit claims rose from 252,000 to 2.9 million, while traditional forecasts indicated an average of around 327,000. However, JoblessBERT was able to predict 2.66 million claims just two days before the end of the reporting week.
On average, over a two-week forecasting horizon, the accuracy of JoblessBERT exceeded traditional methods by 54% at the national level and by 36% at the state level. Additionally, the system can operate at the level of individual cities and fill gaps in official statistics where data is published irregularly.
The researchers also noted the limitations of their approach: their work was based solely on English-language publications, and since 2023, access to Twitter data has become significantly restricted. Nevertheless, they believe that the method can be adapted for other languages and platforms, which is particularly relevant for developing countries with limited capabilities for collecting labor market data.
