FireCast is a pioneering machine learning algorithm project developed in-house at FDNY. The approach screens 7,500 potential risk factors from across NYC government to find proxies that can help predict fires before they occur.
Fires are seemingly random, but there are clues as to why they happen and when to anticipate them. When visiting the scene of a recent incident, there may be attributes that would seem to be obvious precursors of a fire, but why wait for a fire to happen? FireCast was designed for emergency preparedness, that is, providing the field intelligence to fire companies to conduct building inspections. At the heart of it, FireCast is built on a basic spec:
Prob(Fire) = func(human activity, building characteristics, violation activity)
The rest of the algorithm is focused on geography-specific variable selection (e.g. battalions) combined with experimental design to prevent overfitting. As new risk trends emerge, FireCast is designed to capture them and reprioritize buildings for inspection. In short, billions of data elements captured in thousands of variables were tested in thousands of machine learning models tailored to each of FDNY's 49 battalions, allowing FDNY to pinpoint buildings of the highest risk of fire and ensure a safer New York through algorithmic awesomeness.