Which of the following cross validation technique is best suited for time series data ?
MH SET - 2019, SET Exam - 2019,
Answer: D
For time series the right method is forward chaining cross validation.
Observations in a time series are ordered and correlated, so a model must never be trained on the future and tested on the past. Forward chaining (rolling origin) trains on the earliest block and tests on the next, then adds that block to the training set and tests on the following one, and so on, always keeping the test data later than the training data.
K-fold, leave-one-out and stratified shuffle split all assign observations to folds without regard to time order. They let information from later observations leak into the training set, which gives over-optimistic estimates of how the model will forecast.
Please log in or register to report this question.