Cross Validation for Time Series

• Explain why k-fold shuffling is invalid for time series
• Use expanding-window splits for honest forecasting
• Handle lags and feature look-ahead in splits

• Shuffled k-fold leaks tomorrow into the training of today, inflating scores that dissolve live.
• Drift is the norm in product metrics — validation must sample several horizons, not one.
• Expanding windows let small datasets reuse data without breaking chronology.
• Regulators and stakeholders need 'how would this have forecasted last quarter?' not 'how would it fit the past generally?'.

A strategy is only trustworthy if you replay history from the past forward — never look ahead. Forecasting CV is exactly that: refit on the past, predict the next unseen slices, roll forward.

• Order rows by time and split the timeline into train/validation boundaries.
• TimeSeriesSplit: n_splits expanding training windows with fixed or growing test widths.
• Each split: train on history up to boundary, validate on the following window — never the reverse.
• Fit any lags/rolling features on training history only, so their statistics are past-only at validation.
• Average validation scores across horizons → a forecast-backtest estimate of generalization.

• A demand planner validates a forecast model on the last 6 weeks rolling origin: retrain weekly, predict the next week, stack errors.
• A churn team uses time-based splits because viral waves make historical patterns shift.
• An MLOps pipeline reports per-horizon backtest metrics (h+1, h+7, h+30) so freshness targets are explicit.

• Always split by a timestamp, using strictly-future validation windows.
• Train lag features / rolling windows / scaling on each training slice only.
• Report per-horizon errors; a model great at h+1 may collapse at h+30.
• Prefer walk-forward (rolling origin) evaluations for the most realistic coverage.

• Shuffling the time column and 'untimately-valid' leaking day-ahead data into training.
• Fitting rolling means/scalers on the whole series before splitting.
• Using TimeSeriesSplit without gap: train and validation adjacent, so short-horizon lags bleed across the boundary.
• Evaluating on a single horizon and concluding the forecast generalizes.

• Plot the train/validate timeline per split — visual confirmation beats argument.
• If validation beats all history for 'no reason', check lag features computed on data after the boundary.
• Compute the trivial last-value or seasonal-naive baseline in the same splits; the model must beat it.
• Compare h+1 vs h+7 curves: divergence reveals where the model runs out of signal.