Model card

Data coverage

Every source used to build this model, with its refresh cadence, latest verification, and row count. A “verified” status means the source responded and returned rows above the minimum threshold.

NYC-region filter (region 71): kept 2,459 stations, discarded 0.

Architecture

Serving mode: in_process_onnx. Uncertainty method: bootstrap. The runtime scenario evaluator combines precomputed model station features (production/attraction rates, neighbor graph, time-slice features) with a distance-decayed reallocation of pairwise trips across the neighbors on the fly, so a scenario evaluation completes in under 5 seconds.

Trip data may have advanced beyond the training-window end month via public S3 releases since this model was trained. Retraining to incorporate newer months is listed in the “Inside the firewall” roadmap.

Evaluation

Evaluated against 132 accepted validation events. See the Validation lab for per-event residuals and rejection reasons.

Leakage guarantees

Three assertions this model run enforces at training time:

• Not assertedNo post-event trip data in pre-event training rows

  For every accepted validation event, the training set contains no rows with start_date >= event.pre_end. Preserves temporal honesty.

• Not assertedNo current station identity when predicting pre-event

  For every accepted validation event, X_train does not contain the post-event station identity. The model cannot see the answer.

• Not assertedInsufficient status history downgrades confidence, not silently zeros it

  For every station with < 90 days of covered status history, risk_confidence is inflated to 'insufficient_status_history'. The forecast is not shown as if it were reliable.

Known gaps

1. Rebalancing truck logs are internal to Lyft.

   The product infers rebalancing burden from inventory dynamics, not from observed truck routes. Rebalancing truck movements are not observed and would attenuate these effects in practice.

2. Suppressed demand (app searches without a trip) is internal to Lyft.

   App searches that do not result in a completed trip are not in the public data. Suppressed demand is not observed here.

3. 2019-05 to 2023-06 station-status archive gap.

   No public station-status archive covers 2019-05 to 2023-06. Empty/full risk is trained on the OpenBus 2015-2019 window and live GBFS polling from ingest go-live forward. No continuity claim is made across the gap.

4. Pricing and promotions history is unobserved.

   Pricing and promotions history is not observed. It is a confounder that may explain some observed demand shifts.

Inside the firewall

What would materially improve if this workspace were connected to internal Citi Bike / Lyft operations feeds.

Unresolved station IDs

955 stations of the historical set could not be tied to a current canonical identity with ≥ 85% confidence (29.8% of all stations ever observed). They are excluded from training and simulation.