Prediction of the evolution of a claim is a challenging problem in insurance, especially
for guarantees associated with high volatility of the cost such as third-party insurance.
Identifying, soon after occurrence, the claims that require more attention, is particularly
interesting for the company since it allows to better adapt its response to the specificity
of a claim. With the increase of available data on a claim in order to analyze its severity,
artificial intelligence techniques are a promising direction to deal with this problem (see
also [?])). In this paper, we propose an ensemble method using Neural Networks( [?]))
as an early warning system for predicting the cost which is not directly observed due to
censoring( [?])). The model is fed by informations of various types (such as texts reports
about the circumstances of claims and nature of the damage) obtained at the opening of the
claim. A particular attention is devoted to deal with the unbalanced characteristic of our
data, with minority classes representing 2% of our observations. We combine bagging with
a rebalancing method to improve our results and reduce the variance of the estimator. We
illustrate our methodology on the gravity of the accident.