PRODUCTION SCENARIO
A company forecasts weekly store demand from three years of sales history. The team prepares the data in SageMaker Data Wrangler and uses the Split data transform. Validation error is low, but error on the weeks that follow deployment is far higher.
Which split transform produces a validation estimate that reflects production performance?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: An ordered split so the earliest observations form the training set
An ordered split sends the first observations to training and the last to testing, which matches the direction of prediction and keeps later weeks out of training. A randomized split is reproducible with a fixed seed but still trains on future weeks. A stratified split balances the demand distribution and a split by key stops one store spanning both sets; neither one controls time.