Why a 12-Year-Old Forecasting Paper Still Holds Up

At 2025 KDD, the 2014 EMBERS paper received the applied-data-science test-of-time award. This paper was a work involving 30 co-authors from 8 institutions, dealing with methods to forecast social unrest and various events using public data.

EMBERS was a system that combined 5 machine learning models. Using techniques such as Bayesian classification and logistic regression, it analyzed social media, news, blogs, economic indicators, and satellite imagery to forecast changes such as protests, election outcomes, financial events, and health events across 10 Latin American countries. By the time the paper was published, it had already drawn attention for correctly predicting the surge and subsidence of Brazil's large-scale protests in 2013.

The core contribution lay not in the model performance itself, but in the way of evaluating future warnings. Instead of looking at individual warnings one by one, it aligned the temporal sequence of warnings with the temporal sequence of actual events through bipartite matching, finding the most consistent correspondence and keeping it non-crossing where possible to reflect temporal relationships.

Another important idea was that instead of producing results from multiple algorithms all at once, it set prior probabilities on each model's strengths and costs, continuously updating them with Bayes's rule to fuse warnings. This was an approach that systematically combined the outputs of models with different strengths, so that even with many warnings, the result wouldn't become confusing noise.

Srinivasan believes that this kind of Bayesian reasoning survives better not in today's typical supervised learning, but in problems like reinforcement learning, control theory, and agentic AI or robot-based physical AI, where one must choose actions and update beliefs whenever circumstances change. He explains that the same logic remains useful in scientific automation as well, where one selects new experiments and changes subsequent choices based on the results.