How accurate Wolfline arrival times are.

How it's measured

Accuracy

Each time a bus arrives, Dwell looks back at the prediction it was showing minutes earlier, sampled across the trip and bucketed by time-to-arrival exactly as the open-source Transit/Swiftly ETA Accuracy Benchmark specifies (0–3, 3–6, 6–10 and 10–15 minutes out). A prediction counts as accurate if it landed inside that bucket's asymmetric window, stricter on early arrivals, because a bus that has already left is one you can't catch. The headline figure is the benchmark's straight average of the four buckets, built from raw GPS and refreshed hourly.

Completeness

An accuracy score can always be raised by publishing fewer predictions, so it is reported alongside completeness (whether a prediction reached riders at all), measured against the open ETA Completeness Benchmark. Dwell withholds an ETA when the signal behind it has gone stale or a trip can't be matched to a vehicle. Completeness shows what that caution costs.

The two figures above are component rates reported separately, not the benchmark's single completeness score: that score also requires a feed to distinguish a cancelled trip from an untracked one, which Dwell's engine does not yet publish. Until it does, the two parts it does measure are reported as they are.

Where the numbers come from

Nothing here is hand-entered. Both rollups are computed by the engine from its own recorded history and served as public JSON, the same files this page reads, documented on the developers page. The chain that produces them, from a bus's GPS to the minute on a rider's screen, is laid out in how it works.

Wolfline · measurement window since 2026-07-18