Use of a Temporal Test Set

The ultimate test of an empirical model would be to test its ability to predict future data. After all, that is how we generally want to use the model, to predict future data. This simple philosophy makes the definition of test sets by random selection from training sets, or sampled from the training set using ever more sophisticated methods of clustering, rather inconsequential. 

The temporal predictivity of 3 endpoints (logDy 4, solubility in pH 7.4 buffer and human plasma protein binding) with 3 different statistical learners partial 

The quality of a model may change with time and different model learners may show different abilities to extrapolate forward with time. As David Clarke recently highlighted, There is a growing awareness that the validation of ADMET models should not be a one-off process and that models need to be revalidated periodically to make sure their predictive power is not waning as new chemical classes emerge for prediction 

---