NYC Taxi Geo-Temporal Tests

These experiments use 2024 NYC TLC yellow taxi data to test five concrete geo-temporal modeling questions. Results apply to the documented sample and splits; they do not establish universal superiority or evaluate deep models.

Data And Outputs

Path C uses the maintained 2024 yellow taxi benchmark frame from scripts/run_nyc_taxi_quality_benchmarks.py and writes claim artifacts under docs/assets/nyc_taxi_benchmarks/.

Field	Value
Source	NYC TLC trip records
Taxi type	Yellow
Period	2024
Tasks	Duration, fare, pickup demand
Claim runner	scripts/run_nyc_taxi_path_c_claims.py
Result checker	scripts/check_nyc_taxi_path_c_gates.py
JSON artifact	docs/assets/nyc_taxi_benchmarks/path_c_claims.json
JSONL rows	docs/assets/nyc_taxi_benchmarks/path_c_claims.jsonl
Markdown artifact	docs/assets/nyc_taxi_benchmarks/path_c_claims.md

Reproduce

PYTHONPATH=python uv run --group dev --group bench python \
  scripts/run_nyc_taxi_path_c_claims.py \
  --no-download \
  --sample-size 100000 \
  --months 1,2,3,4,5,6,7,8,9,10,11,12 \
  --n-estimators 12 \
  --output-dir docs/assets/nyc_taxi_benchmarks

uv run --group dev python scripts/check_nyc_taxi_path_c_gates.py

The runner stops when real TLC inputs or required taxi-zone geometry are missing. Synthetic smoke output is for development only and is not included in these real-data results.

Questions And Success Criteria

Question	Prediction unit	Evaluation split	Baselines	Success criterion
Directional structure	Pickup zone to dropoff zone by time bucket	Pickup-zone spatial holdout	Unordered pair, source+target additive	Ordered model has at least 2% lower duration and fare RMSE.
Temporal structure	Pickup zone x real hourly timestamp	Rolling-origin zone-time by actual pickup timestamp	Trailing mean, seasonal naive, pooled Ridge	Temporal model has at least 2% lower pickup-demand RMSE than seasonal naive. Trailing mean and pooled Ridge remain reported falsifier rows.
Known-future sensitivity	Zone-time demand	Rolling-origin zone-time by actual pickup timestamp	Future-known covariates ablated	Full model has positive ablation delta and at least 1% lower RMSE.
Spatial transfer	Completed trips from held-out pickup zones	Pickup-zone spatial holdout	Target-encoded zone-only, mean	Trip-level route/geometry features improve duration and fare RMSE by at least 1% over trip-level falsifiers.
Residual correction	Completed trip with baseline estimate	Pickup-zone spatial holdout	Raw baseline, global residual mean, linear residual model	Nonlinear correction has at least 1% lower duration and fare RMSE.

Each result row records the question id, task, split kind, train/test index hashes, dataset hash, model, architecture, capability tier, comparison baseline, primary metric, nonzero materiality threshold, percent improvement, prediction-unit metadata, rolling-origin cutoff timestamp where applicable, RMSE, MAE, WAPE, R2, fit and predict time, peak memory, save/load parity, feature access policy, train-only target encoding, and the guarantee that selection does not use outer test labels.