forecasting-ch13-retail-mllisted

# Chapter 13: The Walmart War Room ## Scope and intake Use when related retail series and known calendar, price or promotion features may support a pooled tree forecast. Do not use for claiming a rolling one-day-ahead exercise is a fixed-origin multiweek forecast, or interpreting SHAP as causal lift. Ask only for unresolved material inputs: Is the decision daily one-step replenishment or a fixed-origin horizon? Were promotions and prices known then? Are sales censored by stockouts? What defines a series? ## Input contract and additional evidence A panel of at least four related series with any number of numeric covariate columns. Declare the covariates to use in config `covariates` and, within them, those whose future values are genuinely known at the forecast origin in `known_in_advance` (a promotion calendar, a list price). Covariates not known in advance are lagged one period by the tool; they never enter at their own timestamp. Minimal **format illustration**, not sufficient training data: ```csv series_id,timestamp,target,promo,price item-0,2021-01-01,53.8,0,10 item-0,2021-01-02,55.1,1,8 item-1,2021-01-01,41.2,0,10 ``` ## Executable interface Exact CLI columns: `series_id,timestamp,target[,covariates...]`. Supported method controls: `ablation, as_of, covariates, frequency, horizon, known_in_advance, lags, origins, rolling, season, seed, shap_rows, strategy`. LightGBM on leakage-safe grouped features: lags (`lags`), shifted rolling means (`rolling`), a calendar