Daily source updates, not live streaming. The newest partial dates are excluded. Estimates between the data cutoff and today are nowcasts. Capacity changes the scenario calculation, not the forecast model.
Observed demand & forecast
Where the forecast is weakest
Do the uncertainty bands earn their coverage?
The capacity trade-off
Retrospective performance · separate from live results
Published forecast performance
| Target week | Predictions | Distinct dates | WAPE | Band coverage |
|---|
Results by published model version
| Version | Scored | Dates | Coverage | WAPE | Pending |
|---|
Coverage investigation and calibration experiment
Only forecasts published before the target day enter this chart. Actuals mature for seven days after the source cutoff passes their date. Overlapping forecasts are counted separately; this is not a sample of independent outcomes. Coverage warnings require at least 28 forecasts over 14 target dates.
Why this forecast looks this way
| Candidate | Window 1 WAPE | Window 2 WAPE | Window 3 WAPE | Pooled WAPE |
|---|
Selection uses MAE; WAPE has the same ranking within each shared evaluation period. Equal-length windows contribute equally to pooled MAE. The later test scores do not select the winner. These are retrospective comparisons; future published results are the check on this revised policy.
Demand can change because of weather, holidays, reporting access or data revisions. These are possible explanations, not established causes. A 28-day demand shift above 30% is flagged for investigation. NYC reporting changes · Weather attribution: Open-Meteo / ERA5
Daily planning view
| Date | Expected | Lower | Upper | Capacity gap |
|---|
The capacity scenario assumes a fixed number of requests processed each day. It excludes existing backlog, case complexity, staffing availability, and carryover; it is not a staffing recommendation.
How to interpret this forecast
Promotion and rollback policy: reviewed releases, four weekly challenger wins, and an explicit coverage escalation rule. Automatic model fallback is not enabled.
- Target: recorded requests created per calendar date, using the source's New York local timestamps. Only aggregate counts are downloaded; addresses and request descriptions are not used.
- Models: last available matching weekday, mean of four matching weekdays, and regularized seasonal regression using past demand and calendar features. Candidates must improve pooled validation MAE by at least 2% and win at least two of three windows to replace their reference baseline.
- Error bands use a separate 42-day calibration period. Nominal 90% coverage is an empirical target, not a guarantee. The observed test coverage is shown above.
- Current source records can be corrected retrospectively. Backtests use the current extract, not archived historical publication vintages. Past data-publication delays are not reconstructed.
- Borough and city forecasts are fitted separately and need not add up. City totals include requests with unknown borough. Requests do not measure all community needs or service quality.
- Unexpected events, reporting changes, holidays, and weather can shift demand. No causal or operational-performance claims are made.