Arrival — ETA prediction with a point-in-time feature store Diagram

Why two paths, one declaration

A feature has a value only as of a moment in time
Training asks as of each trip's assigned_at; serving asks as of now
test_parity replays 600 trips asserting the two agree exactly

What a leak costs

Honest model beats the flat promise by 2.830 min MAE
Filtering the window on assigned_at looks 0.924 min better
Dropping the window gaps only 0.086 min — the dangerous kind

The blind window

A trip's duration is known ~35 min after it was predicted
20 min into an outage, MAE reads identical to a healthy model
Prediction drift moves immediately: PSI 10.3 against 0.0