RESEARCH · RESEARCH · #1648
Layered diagnostic study finds better traffic forecasts do not necessarily improve signal control
This arXiv preprint presents a layered diagnostic framework using 29 days of reconstructed demand from Xuancheng, China, to trace why improved entry- and movement-level forecasts (≈4% MAE reduction) and calibrated conformal intervals often fail to yield better closed-loop signal-control decisions. The authors identify interface leakage, limited effective action choices at intersections, and misaligned objectives as key failure modes, show a synthetic positive control where future information reduces rollout cost by 61.5%, but find that on frozen test dates causal forecasts and an event oracle increased queue vehicle-seconds versus a matched no-future rollout.
KEY POINTS
- This arXiv preprint presents a layered diagnostic framework using 29 days of reconstructed demand from Xuancheng, China, to trace why improved entry- and movement-level forecasts (≈4% MAE reduction) and calibrated conformal intervals often fail to yield better closed-loop signal-control decisions.
- The authors identify interface leakage, limited effective action choices at intersections, and misaligned objectives as key failure modes, show a synthetic positive control where future information reduces rollout cost by 61.5%, but find that on frozen test dates causal forecasts and an event oracle increased queue vehicle-seconds versus a matched no-future rollout.
- This matters because it shows that gains in predictive accuracy can fail to produce operational benefit in control systems and provides a practical protocol to diagnose where and why forecast improvements do not translate into better decisions.
WHY IT MATTERS
This matters because it shows that gains in predictive accuracy can fail to produce operational benefit in control systems and provides a practical protocol to diagnose where and why forecast improvements do not translate into better decisions.