Learning Calibrated Corrections to Spend-Based Carbon Estimates for ICT Hardware

Accurate, actionable, and scalable carbon estimation methods are required for developing emissions reduction strategies for information communication technology (ICT) hardware. Component-level estimates are aggregated to a system and fleet level which is then used in the pursuit of high-value carbon reduction efforts. Spend-based estimation methods are broadly applied across the industry yet suffer from two notable drawbacks. Spend-based emissions factors are calibrated using a top-down approach at a single point in time and therefore generate estimates that are vulnerable to market price shifts and lack the resolution to distinguish differences within a sector. This proposal covers a framework for learning spend-based correction factors using conformalized quantile regression. The resulting calibrated intervals are monitored to detect drift and trigger recalibration as inventories evolve, keeping the corrections interpretable over time.