Mark Last

The warranty datasets available for various car models are characterized by extremely imbalanced classes, where a very low amount of under-warranty vehicles have at least one matching claim ("failure") of a given type. The failure probability estimation becomes even more complex in the presence of censored warranty data, where some of the vehicles have not reached yet the upper limit of the predicted interval. The actual mileage rate of under-warranty vehicles is another source of uncertainty in warranty datasets. In this paper, we present a new, continuous-time methodology for failure probability estimation from multi-dimensional censored datasets in automotive industry.