After a number of Google searches and looking at various stack posts, I cannot find much information or discussion about using stratified cross-validation in regression context.
I am modeling forest biomass using satellite data. One challenge is that the distribution of the response has a negative skew. This is because old growth forests with large biomass densities are quite rare on the landscape but still important to include in the model (they have considerable ecological importance). Conversely, samples with low biomass densities are quite common.
In general, when evaluating a machine learning model, one would like each fold to be representative of the entire training set to approximate the generalization error. It seems to be common, in a classification context, to use stratified folds. However, I can find almost no discussion of using stratified folds in a regression context. Is there a reason for this?
My hunch is that there doesn't seem to be a straight forward way to perform the stratification (perhaps though binning)?