Using Neural Networks to Predict Claim Duration in the Presence of Right Censoring and Covariates

Data Mining & Neutral Networks (narrow topic or advanced); We present a general methodology for fitting feed-forward neural networks when both right censoring and covariate information (claim attributes) exist. Right censoring occurs when only intermediate, but not final values of a time-dependent variable (such as claim duration) are known for some data points, and final values of the variable are known for all other observations. This situation frequently arises in casualty insurance when there are active claims in an analysis data set. The techniques we develop are applicable for estimating the distribution of claim lifetimes when awards are disbursed over the unknown claim life. The neural-network framework allows us to handle complex relationships between the claim attributes and claim duration. We will derive a generalization for right-censored data of the back-propagation method used for fitting feed-forward neural networks. A connection between least squares estimation and maximum likelihood estimation will be used to establish the generalization. A typical cross-validation approach to modeling will be described to reduce over-fitting. An application of our methods is demonstrated for predicting the duration of a claim in worker's compensation insurance in the presence of covariates.