DiDi's EXHOLD System Optimizes Ride-Hailing Matching Experience

In large-scale ride-hailing platforms, 'hold control' is a crucial mechanism that strategically delays certain driver-order matches to wait for better opportunities. This helps reduce cancellations and minimizes wasted effort for drivers. However, existing industrial hold strategies often rely on brittle heuristic thresholds derived from multiple predictive models, making them difficult to optimize for complex, multi-objective experience signals in dynamic traffic conditions. DiDi has introduced EXHOLD, a deployable two-stage framework designed to address these limitations. The first stage involves a decision model that assigns each driver-order pair to discrete, interpretable experience tiers. This model is optimized using a unified objective that aggregates satisfaction signals across the entire matching funnel, from initial request to trip completion. The second stage of EXHOLD solves for a monotone hold-time schedule through constrained optimization over empirical quantiles. This approach explicitly enforces service guardrails, preventing unnecessary delays for promising matches while maximizing overall experience improvement. Randomized A/B experiments in DiDi's production system in Brazil demonstrated consistent gains in marketplace efficiency and experience, including increased trip completion, higher driver income, and significantly reduced passenger cancellations. EXHOLD is now fully deployed and serving production traffic.