Zodiac — Zero-inflated overshoot controlled dual-head integration for asymmetric cross-domain forecasting

Foundation models promise zero-shot forecasting across domains, yet their effectiveness for cold-start scenarios with zero-inflated distributions remains underexplored. We study cross-domain demand forecasting, predicting outcomes for items launching in new domains without historical data where a substantial fraction of launches (≈ 30%) yield zero outcomes and overestimation carries asymmetric costs. We propose a specialized architecture—ZODIAC—combining: (1) dual-domain temporal integration via stacked recurrent layers processing source and target domain signals, (2) a dual-head design with classifier and regressor explicitly modeling zero-inflated distributions, and (3) an asymmetric loss function penalizing overestimation to align with domain-specific costs. We benchmark our approach against a pretrained in-context learner (TabPFN), an AutoML ensemble (AutoGluon), and a neural time-series model (Temporal Fusion Transformer) across six cross-domain forecasting tasks. Our model achieves 80% WAPE, a 13% relative improvement over TabPFN, 25% over AutoGluon, and 26% over TFT while reducing systematic overprediction from 66–87% to under 41%, a property unachievable with models lacking loss customization.