The system, named JoblessBERT, is trained to analyze Twitter (X) posts where users report job losses. Unlike traditional methods that look for specific phrases, this neural network is capable of recognizing colloquial expressions, slang, and even typos, allowing it to detect 13 times more job loss cases with high accuracy.
To account for the biases of the Twitter audience, the researchers identified users' age, gender, and geographical location based on their profiles and adjusted the data according to census results.
The model was tested on data from 2020 to 2022, including the pandemic period. For example, in March 2020, when unemployment claims surged from 252,000 to 2.9 million, traditional forecasts estimated around 327,000, while JoblessBERT predicted 2.66 million just two days before the end of the reporting week.
On average, the accuracy of JoblessBERT's predictions over a two-week horizon was 54% higher than that of traditional methods at the national level, and 36% better at the state level. This system can also be used to analyze data at the city level and fill gaps in statistics where data is published irregularly.
The authors emphasize the limitations of their approach: the study was conducted only on English-language posts, and since 2023, access to Twitter data has become more restricted. Nevertheless, they are confident that the method can be adapted for other social platforms and languages, which is particularly relevant for developing countries with insufficient access to labor market data.
