A forecasting model is useful only when its inputs existed before the outcome it predicts. Random row splits can hide that constraint by mixing later observations into training. You will hold out a chronological suffix, inspect expanding validation boundaries and compare a future forecast with a last-observation baseline. The first validation row is already outside training; an off-by-one index can leak it back into the fit. A gap can represent a delay in label availability. The data in this lesson is a fictional order sequence with trend and repeated variation, so a future error answers a different question from a random-sample error.