Global tree forecasters collapse at the hierarchical aggregate: a scale-aware, library-agnostic correction

Abstract

Global gradient-boosted-tree forecasters, trained by pooling series, are a workhorse of demand forecasting and reported accurate across hierarchy levels. We document a deployment-blocking exception. On a production panel of monthly gross transaction value from a business-to-business service marketplace, a global tree scored at the marketplace aggregate—far outside its per-buyer training support—collapses, under-predicting the total by an order of magnitude. We trace this to the constant extrapolation of trees beyond their training range: the failure is fundamentally one of scale, curable along either of two axes. One restores scale, via per-series scaling or a single-model cohort-weighted aggregate training row, both library-agnostic. The other renders the target stationary by seasonal differencing, which uniquely remains stable under recursive multi-step forecasting. The failure is structural—reproducing on synthetic data and three public datasets (M5, Tourism, B2B)—and seed-invariant, situated within a twenty-three-method benchmark documenting a unit-of-analysis dichotomy.