Build a leakage-free backtest for a weekly retail forecasting model

Your team reports WAPE 0.18 for a new LightGBM demand model validated with random 5-fold cross-validation. The incumbent seasonal-naive-plus-ETS blend is at 0.24 in production reporting. Before anyone celebrates, you are asked to design the evaluation that the finance director can trust.

1. List the leakage routes in this setup, from splits to features to covariates.
2. Design the backtest: origins, gap, horizons, the series set, and metrics.
3. Which numbers do you report, and how do you aggregate over millions of series?
4. How would you compute forecast value added, and for what steps of the pipeline?
5. In your own experiments random k-fold was not optimistic. Should you therefore keep it? Explain.