39e4973ba3321b80f37d9b55f63ed8b8-AuthorFeedback.pdf

We herein use the median t i+1 of the predictive distribution to predict ti+1 for each5 model. To obtain the median t i+1, we use the relation Φ(t i+1 ti|hi) = log(2), where Φ is a cumulative hazard6 function. This relation is derived from the property that the integral of the intensity function over [ti,ti+1] follows7 the exponential distribution with mean 1. Then, the median t i+1 can be efficiently obtained by solving the relation8 using a root finding method (e.g., bisection method); it takes only a second for our model to generate predictions for9 20000 events. Therefore, the cumulative hazard function also plays a crucial role in generating a median predictor.10