A comparison of three strategies for forecasting warranty claims

The economic necessity for developing accurate forecasts of extended warranty costs on durable goods is established. Three strategies for forecasting warranty failures are outlined. The first strategy encompasses the use of linear, predictive models, wherein forecasts are based upon an extrapolation of a least squares fit of the data. These include regression and time series approaches. The risks associated with the use of these models when localized or highly nonlinear phenomena exists are illustrated. The second strategy is based upon the use of dynamic linear models wherein the model parameter estimates are updated sequentially by using the equations for the Kalman filter. The adaptive nature of these models leads to improved forecasts in the presence of localized phenomena. The third strategy encompasses the use of nonparametric modeling approaches, including the use of neural network models to provide a generalized, form-free modeling framework. With a constructed example, the comparative advantage in the use of a neural network representation is demonstrated through the use of a cross-validation strategy to identify and compare the best model in each class of models under consideration.