RESEARCH · RESEARCH · #1647
Empirical evaluation of joint upper‑bound coverage for route choice (arXiv:2610.06995v1)
This arXiv preprint evaluates whether improving joint upper‑bound coverage of path travel‑time forecasts helps downstream route decisions. Using offline proxy tasks built from Beijing and Chengdu road speed data (150 OD pairs, three candidate paths, 14 test days), the authors compare raw 90th‑percentile path time bounds to jointly calibrated upper bounds: joint calibration raises joint coverage substantially (e.g., from 83.19% to 92.26% in one Beijing task) but nevertheless increases lateness and mean travel time in the frozen tasks; separately, a predictor with 14.91% lower speed MAE reduced lateness modestly under a different decision rule. The paper stresses that joint coverage is not a reliable surrogate for route utility in these offline experiments and does not claim online or causal benefits.
KEY POINTS
- This arXiv preprint evaluates whether improving joint upper‑bound coverage of path travel‑time forecasts helps downstream route decisions.
- Using offline proxy tasks built from Beijing and Chengdu road speed data (150 OD pairs, three candidate paths, 14 test days), the authors compare raw 90th‑percentile path time bounds to jointly calibrated upper bounds: joint calibration raises joint coverage substantially (e.g., from 83.19% to 92.26% in one Beijing task) but nevertheless increases lateness and mean travel time in the frozen tasks; separately, a predictor with 14.91% lower speed MAE reduced lateness modestly under a different decision rule.
- The paper stresses that joint coverage is not a reliable surrogate for route utility in these offline experiments and does not claim online or causal benefits.
WHY IT MATTERS
It shows that better probabilistic coverage (calibration) does not necessarily translate into better route‑choice outcomes, cautioning researchers and practitioners against using coverage as a proxy for decision utility.