Causal forecasts raise queue vehicle-seconds 6.09% on Xuancheng test dates despite 4.03% MAE reduction

The study reconstructed 29 days of demand at nine intersections and tested point forecasts, conformal intervals, and dependence-aware scenarios against closed-loop controllers. Entry-level forecasts cut MAE 4.03% versus historical means, yet nominal 90% intervals covered only 75.66% of ex-post high-demand periods. Interface audits exposed decision-time leakage and showed that only two intersections possessed multiple effective actions under the deployed controller.

On the held-out dates, both causal forecasts and a five-second event oracle raised internal rollout cost while the oracle trimmed spillback exposure 3.78%. Paired-day bootstrap intervals crossed zero, indicating no reliable operational gain. A synthetic positive control confirmed that perfect future information could cut cost 61.5%, isolating the gap to temporal observability, action identifiability, and objective misalignment.

The layered protocol quantifies where predictive accuracy fails to propagate: forecast error, coverage on tails, leakage at the decision interface, and mismatch between queue-vehicle-seconds and spillback penalties. These diagnostics apply directly to any forecast-to-control pipeline where action sets are sparse and dynamics are partially observed.

Subsequent deployments should instrument action identifiability and objective gradients before scaling forecast resolution.