Cross-validation for dependent data
Harvard Rue (KAUST)
Tuesday 3rd October 16:00-17:00 Joseph Black A504
I will discuss our new take on cross-validation (CV) for dependent data. Traditional use of CV, like leave-one-out CV, is justified using independence-like assumptions. With dependent data, leave-one-out CV makes less sense, as we are evaluating interpolation properties rather than prediction properties. We can adapt the CV idea to dependent data by removing a set of "near-by" data-points (to be defined) before predicting, but the issue is then how to do this in practice, which is less evident for more involved models. I will discuss our approach in the context of Latent Gaussian Models (LGM) where we can automatically select appropriate groups of data to remove before predicting one data point. The new group-CV approach is available in the R-INLA package.