COCOA: a model-agnostic framework for forecasting accessories demand from final product sales

Forecasting accessories demand is challenging because it is structurally dependent on final product sales but not deterministically derived from them. Existing approaches either treat accessories as independent time series or rely on attach-rate assumptions embedded in model-specific frameworks, limiting flexibility and generalizability. This paper proposes the COCOA (Co-Occurrence and COvariate-based forecasting of Accessories) framework, a model-agnostic framework that exploits transaction-level co-occurrence and integrates forecasted final-product demand as future covariates within a multivariate model. The approach is evaluated through a real automotive case study involving 121 accessories with heterogeneous demand patterns, benchmarked against univariate models across multiple forecasting techniques and horizons, under both perfect and imperfect final-product forecasts. Results show significant and largely model-agnostic improvements for intermittent and lumpy demand, with gains attenuating but remaining present under realistic forecast uncertainty.