Hi, I have created simple random split, and cv is inline with lb, however, I did not create any lag features, in which case this split is not meaningful. In that case, keeping the last few weeks of data for validation should be correct approach or use timeseries cross validation.