Taxi Zone Feature Checks

Bottom Line

This benchmark checks whether CartoBoost can express taxi-lane feature families: sparse lane membership, route midpoint geometry, periodic hour, and combined geotemporal effects. It is deterministic feature acceptance, not a model comparison against external baselines.

Reproduce

uv run python scripts/run_lane_level_acceptance_metrics.py \
  --output-dir docs/assets/lane_level_tests

Artifacts:

• docs/assets/lane_level_tests/acceptance_metrics.json
• docs/assets/lane_level_tests/acceptance_metrics.md
• docs/assets/lane_level_tests/lane_heatmap.png
• docs/assets/lane_level_tests/hour_profile.png
• docs/assets/lane_level_tests/route_midpoint_geometry.png

Results

Feature check	Main result	Outcome
Sparse lane membership	Sparse lane RMSE 0.0; axis lane-ID RMSE 51.96	Passed: exact sparse lane fit.
Hot lane isolation	Hot lane margin 240.0	Passed: margin > 200.
Route midpoint cartometry	Gaussian midpoint RMSE 0.0; axis midpoint RMSE 69.40	Passed: Gaussian route RMSE < half axis baseline.
Periodic hour wraparound	Periodic hour RMSE ~0; axis hour RMSE 31.58	Passed: 23:00 and 01:00 edge gap 0.0.
Combined lane/spatial/temporal	Holdout RMSE drops from 12.68 to 4.86	Passed: full model / axis ratio 0.383.

Plots

Interpretation

These results show that the implementation can represent each targeted feature family on a controlled taxi-lane fixture. They are useful for diagnosing model behavior before running a real-data benchmark.

Passing does not mean CartoBoost is more accurate than LightGBM, XGBoost, or a forecasting library on real data.