0.00.60 alert0.851.0station with no score in this window
Live gaps use the first snapshot of two distinct upcoming trips at the same route, stop and direction. Departure-only updates are excluded. Repeated trip pairs are suppressed within a four-hour window; this cache resets on restart. Feed freshness is separate from the age of the latest scored pair. Scores describe deviation in arrival estimates; they are not measured train passages, calibrated probabilities or official MTA incident reports.
Counters describe the live-source window; replay has separate measurements.
—
stops scored
—
rows scored
—
anomalies ≥ 0.60
—
high score ≥ 0.85
—
newest data
Observation queue
Loading live scores…
Pipeline
Checking…
Continuous collection / chronological evaluation
History & forecasting
Loading history…Waiting for a validated worker snapshot
Follow the evidence as the dataset grows. The target is a future feed-predicted arrival-spacing proxy, not actual train passages or passenger waiting time.
Retained history span
—
Complete fresh windows
—
Five-minute windows
—
Stable route/direction groups
—
Forecasts appear only after history, coverage and freshness checks pass.
Route
Platform direction
Horizon
Predicted spacing
Algorithm
Target time
Inspect temporal evaluation & collection health
No measured temporal evaluation is available yet.
Horizon
Method
MAE
RMSE
Measured / eligible pairs
Coverage
Lower error is better. Missing 24-hour seasonal values mean no comparable lag was available. Validation-selected rows can combine different algorithms across route/direction groups; the test set does not choose the method. Platform N/S uses the documented MTA stop suffix, not an inferred compass bearing.
Measured on the frozen labelled replay(216 rows, 16 labelled, 3 incidents, 6 stations — small by design):
recall@20 1.00 vs 0.625 for the best of three baselines, 3/3 incidents caught,
6.7 min average lead time —
at the highest false-alarm rate of the four (0.030 vs 0.000–0.010).
Full table below.
online model (River)rolling z-scoreEWMAfixed thresholdlabelled incident windowalert threshold 0.60
Evaluation & method notes Inspect the measured trade-offs
Replay evaluation — every number below comes from the frozen artifact
MTA Scan