I pulled 18 months of stop-level scans from two parcel carriers into BigQuery and trained an XGBoost model, but MAE is stuck around 0.9 days on Northeast lanes during peak — has anyone improved drift using features like NOAA weather, facility throughput, or geofenced dwell times? Goal is to shift alerts to 2-hour buckets without hammering ops with false positives.
But > dwell times? Goal is to shift alerts to 2-hour buckets without hammering ops with false Agree — what moved the needle for us was a facility-hour dwell delta feature (actual minus 28‑day median) fed into lane-level quantile XGBoost with week-of-year residual calibration; that pushed NE peak MAE to about 0.4d and made 2‑hr buckets usable — can you compute that delta in BigQuery and back‑propagate it along the linehaul plan?
I’d keep your BigQuery + XGBoost but add a residual layer: per lane-week quantile regressor on the XGB error using NOAA hourly precip/wind along the route and a facility throughput z‑score; that cut our NE peak MAE from about 0.9d to about 0.6d and made “2‑hour buckets” workable. If you can, build a 48h backlog index = inbound scans / planned capacity per facility and lag it 6–12h; @rbaker’s dwell delta helps, but the backlog signal caught Friday surges.
What helped us was an “origin cut‑time slip” feature (minutes past scheduled trailer close) plus a lane×carrier×hour bias term updated daily via EWMA on recent errors — a little Kalman‑ish nudge. We also wrapped XGB with conformal calibration to keep alert bands calibrated and used NWS CAP alerts over raw precip: https://api.weather.gov/alerts. @wenchen have you tried a per‑lane bias tracker instead of refitting a residual model?
Add a ‘scan‑gap velocity’ feature — minutes since last movement scan normalized by the 7‑day median for that facility×hour — and then conformalize the XGB residuals per lane with a sliding window to get calibrated bucket probabilities. It cut false pings for us without dulling sensitivity, like swapping a guess for a fuel gauge. Building on @rlewis: do you have an air/ground route flag or sort‑plan revision ID, since regime flips there were the hidden drift driver for us?