Using out-of-sample Cox-Snell residuals in time-to-event forecasting
The problem of assessing out-of-sample forecasting performance of event-history models is considered. Time-to-event data are usually incomplete because the event of interest can happen outside the period of observation or not happen at all. In this case, only the shortest possible time is observed and the data are right censored. Traditional accuracy measures like mean absolute or mean squared error cannot be applied directly to censored data, because forecasting errors also remain unobserved. Instead of mean error measures, researchers use rank correlation coefficients: concordance indices by Harrell and Uno and Somers’ Delta. These measures characterize not the distance between the actual and predicted values but the agreement between orderings of predicted and observed times-to-event. Hence, they take almost “ideal” values even in presence of substantial forecasting bias. Another drawback of using correlation measures when selecting a forecasting model is undesirable reduction of a forecast to a point estimate of predicted value. It is rarely possible to predict the timing of an event precisely, and it is reasonable to consider the forecast not as a point estimate but as an estimate of the whole distribution of the variable of interest. The article proposes computing Cox–Snell residuals for the test or validation dataset as a complement to rank correlation coefficients in model selection. Cox–Snell residuals for the correctly specified model are known to have unit exponential distribution, and that allows comparison of the observed out-of-sample performance of a forecasting model to the ideal case. The comparison can be done by plotting the estimate of integrated hazard function of residuals or by calculating the Kolmogorov distance between the observed and the ideal distribution of residuals. The proposed approach is illustrated with an example of selecting a forecasting model for the timing of mortgage termination.