Modeling loss and no-loss fire incidents using artificial neural network: Case of Toronto

A predictor neural network was proposed for loss prediction of fire incidents. Such a predictor could help to tackle loss predicted incidents more effectively in order to reduce the number of actual loss incidents. A fully connected multilayer feed-forward neural network was adapted for the prediction task. The network was trained with 8337 fire incident records of the Toronto data set reported between 2000 and 2006, and then its performance was evaluated on 2778 never seen records. The output of the network was interpreted in two different ways: first as a probabilistic prediction and second as a binary prediction. Results obtained reported a very decent ability of this approach to predict a loss fire incident.