Consider "backend GP" for uncertainty estimation on non-GP models

[ardunn]

A GP can be trained on the residuals of predictions in order to estimate uncertainty. This may be superior to bootstrap estimation in terms of computation time at least.

[ardunn]

alternatively, just additionally fitting a gp and using the uncertainty from that.