Omni-channel grocery forecasting: Channel differences in forecastability and predictive signals

Customer behavior in omni-channel grocery retail creates forecasting challenges because shoppers switch between offline and online channels. We assemble transaction-level data from loyalty card membership of a major South Korean grocer and enrich them with competitor and promotional information to evaluate channel-specific sales predictability. Drawing on recent evidence that online grocery shopping generates more inertial purchasing patterns than offline shopping, we develop three propositions that link channel-specific consumer behavior to forecasting design. Using XGBoost with strict out-of-sample tests and SHAP-based interpretation, we document three main findings that support these propositions. First, forecast accuracy increases as features are added, with the largest gains from customer purchase histories. Second, online sales are more predictable than offline sales. Third, predictability is structured by product taxonomy: online sales are best explained by broader categories (classes), whereas offline sales are better explained by narrower subcategories (subclasses). These findings support our propositions that purchase histories matter the most because they most directly encode inertial purchasing patterns, online channels are more predictable because they exhibit higher inter-trip purchasing consistency, and taxonomic asymmetry reflects the level at which purchasing inertia is strongest in each channel. Retailers should plan at the class level online and curate at the subclass level in physical stores to improve assortments, promotions, and waste reduction.