Public methodology register
How late the worst 5 in 100 settled predictions were
- Version
- last-prediction-within-ten-minutes-v1
- ID
- network-reliability:p95-delay
Definition
- Definition
- Sort selected final predicted delays from least to greatest. This is the delay where about 5 predictions in 100 were equally late or later in the published route window.
- Formula
- PERCENTILE_DISC(0.95) ordered by settled delay seconds.
- Units
- seconds, displayed in minutes
- Population
- The same selected final provider predictions used by the on-time percentage.
- Measurement window
- Rolling 30-day route-health window.
- Coverage
- Matched published predictions only; the discrete percentile selects a final predicted delay value.
- Publication eligibility
- Regular comparable bus routes with publishable final provider-prediction evidence; merged route variants with incompatible distributions withhold this percentile.
- Scope
- Each route’s settled final provider predictions in the rolling 30-day route-health window.
- What this number can support
- An upper-tail final-prediction schedule difference: about five selected predictions in one hundred were this late or later. This is not an independently measured arrival.
Implementation status
- Source candidate
- Implemented in this source candidate.
- Public production
- The corrected metric version is implemented in source; public production deployment and verification remain unrecorded.
The final provider-prediction calculation is implemented in this source candidate. Public production verification for this corrected version is still not recorded.
Sources and identifiers
- Method
last-prediction-within-ten-minutes-v1Settled provider-prediction reliability method - Current API
reliability-evidenceCurrent network reliability public API
Exclusions
- Predictions more than ten minutes before the scheduled arrival are excluded, and delays outside ±90 minutes are removed as timetable-wrap outliers.
- OC Transpo rail is excluded because the operator does not publish realtime rail observations in the feed used by this method.
Uncertainty and interpretation
- The result describes matched published predictions, not every operated trip. Feed gaps and trip-matching limits can reduce coverage.
- Current network evidence is recomputed and is not stored here as an immutable historical snapshot.
Current evidence
Change history
No registered material change is recorded for this metric version.