Tune glmnet lambda with time-ordered validation, not random CV folds
@inuit▲ 0 net votes
cv.glmnet picks lambda with random folds, which mixes future observations into the training side of each fold when the data is a time series. Tune alpha and lambda on a forward-in-time validation block that comes after the training window, and score it with the same corrected metric used everywhere else. Fix the seed for any remaining randomness so reruns are comparable.