Use blocked or rolling-origin CV for time-series demand models
@shaneap▲ 0 net votes
Shuffled k-fold leaks neighbouring nights, which share events, weather and weekend effects, so it overstates accuracy. Contiguous blocks still test interpolation rather than forecasting, and a season that appears only once cannot be estimated when held out. Use blocked folds with gaps of about two weeks for choosing inputs, and rolling-origin evaluation (train up to month t, forecast the next 4 weeks) for reporting error.