Design a forecast backtest that respects time and information

Design an evaluation plan for this forecast.

Decision, target and horizon: Forecast daily support-ticket volume seven days ahead to plan staffing; underprediction is more costly than modest overprediction.
Data history and feature availability: Two years of daily counts; weekday and public-holiday flags are available ahead of time. Campaign counts are finalized only after each day.
Candidate model and comparison baseline: Candidate: gradient-boosted regression using lagged counts. Baseline: the count on the same weekday one week earlier.

Specify chronological training and evaluation windows, a rolling-origin schedule, retraining cadence and any gap needed between windows. Explain which information is available at each forecast issue time; flag future-known or revised features that could leak outcomes.

Choose one primary error measure tied to the decision, plus a useful secondary measure and segment checks. Address zeros, scale differences and asymmetric error costs where relevant. Compare against the supplied simple baseline under identical windows.

Include uncertainty evaluation if intervals are produced and a reproducibility checklist. Do not train models, invent accuracy, or choose the winner without actual out-of-sample results.