Live demo: venius24.streamlit.app (pick two zones and a time, get an estimate with an 80% range)
Predicts how long a NYC yellow taxi ride will take using only what is known at pickup (time, pickup and dropoff zone), together with an 80% range for the ride. Trained on April 2026, then tested on months the model never saw: May (v1, ~4M trips) and June (v2, ~3.7M trips), each opened exactly once.
Why it matters: an ETA shown to a rider or used for dispatch is only useful if it holds up on future days. This project is about measuring that honestly: no leaked features, time-ordered validation, frozen test months.
| model | May 2026, MAE / RMSE (min) | June 2026, MAE / RMSE (min) |
|---|---|---|
| global median | 10.01 / 16.05 | 8.97 / 13.51 |
| median by route and hour (all April trips) | 4.92 / 8.35 | 4.43 / 7.05 |
| Ridge (log target, zone geometry) | 5.11 / 8.19 | not run |
| v1: LightGBM, 500k-trip sample | 4.36 / 6.97 | 4.07 / 6.26 |
| v2: LightGBM, all 3.74M trips, bigger trees | not run (May was spent on v1) | 4.01 / 6.16 |
- v2 vs the route-hour median: 9.6% lower MAE and 12.6% lower RMSE on June; better on 28 of 30 days.
- v2 vs v1: 1.6% lower MAE, better on all 30 June days. Small, as the learning curve predicted.
- June is an easier month than May for every model (even the global median), so compare models within a column, not across.
80% interval (v2): on June it covers 80.5% of trips (target 80%) with a mean width of 12 min. Instead of a single number the model returns a range that contains the real duration 8 times out of 10: about 10 min wide for short hops and 20 min for trips over 8 miles. The range comes from two quantile models (P10, P90) calibrated with conformalized quantile regression (CQR); details in docs/methodology.md.
All numbers come from scripts in this repo (reports/final_results.md,
reports/final_results_v2.md); none are typed by hand.
Try it: venius24.streamlit.app. The first load takes about half a minute while the models load.
A small Streamlit app (app/streamlit_app.py) runs the v2 models: pick two zones, a date and a pickup
time, and it shows the estimate, the calibrated 80% range, and how both change across the day.
To run it locally:
pip install -r app/requirements.txt
streamlit run app/streamlit_app.py # downloads the trained models from the GitHub release on first startTo retrain the app models yourself: python scripts/train_app_models.py (about an hour on 12 threads; --quick for a rough
one-minute version).
- No leakage. Trip distance, fare, tips and dropoff time are never features. A guard in code (
assert_no_leakage) and tests enforce it. - Time-ordered validation. 5 expanding-window folds inside April; each fold trains only on trips that ended before validation starts (3-hour gap).
- Each test month is opened once. All choices (model, size, interval calibration) are made by CV first; the evaluation scripts refuse to run twice. The v2 choices were committed and pushed before June was opened.
- Two independent holdouts. v1 was tested on May; v2 on June. v1 was re-run on June too, so v1 vs v2 is compared on identical data.
- Reproducible. Fixed seeds, deterministic LightGBM, pinned versions; one command rebuilds every report.
| step (all CV on April, MAE min, mean ± std over 5 folds) | MAE |
|---|---|
| route-hour median baseline (500k sample) | 5.481 ± 0.134 |
| v1: LightGBM, 500k-trip sample, 127 leaves, 600 trees | 4.203 ± 0.110 |
| same model, 2M trips | 4.139 ± 0.110 |
| same model, all 3.74M trips | 4.124 ± 0.102 |
| v2: all 3.74M trips, 255 leaves, 600 trees | 4.106 ± 0.099 |
More data helps with diminishing returns; larger trees help a little once the data is there. Differences between the top
configurations are smaller than the fold-to-fold spread, so the pre-declared rule "lowest mean CV MAE" picks the winner.
Details: reports/learning_curve.md, reports/cv_v2.md.
- Long trips: ~10 min MAE beyond 8 miles; zones and hour cannot see route choice or traffic. The 80% interval covers only 74% of these on June.
- Airports (JFK, LaGuardia): highway congestion we have no data for.
- Drift: trips got longer over April, so models trained on earlier days under-predict. Weighting recent trips helped a little.
- Holidays: April has none, so on Memorial Day (May) the model gained only 1% over the baseline.
pip install -r requirements.txt
python scripts/download_data.py # ~200 MB of public TLC data (Apr, May, Jun), not stored in the repo
pytest -q # 28 tests
python scripts/run_all.py # EDA, CV, error analysis, both final evaluations, summaryWant a quick look first? notebooks/demo.ipynb trains the final model in ~20 s and asks it about real
routes (Midtown to JFK at rush hour vs 3 am, and more). It renders on GitHub with outputs; to run it yourself,
pip install jupyter and download April only: python scripts/download_data.py 2026-04.
src/taxi/ data cleaning, features, zone geometry, models, CV, prediction intervals
scripts/ one script per step (eda, run_cv, run_cv_v2, error_analysis, final_eval, final_eval_v2, ...)
tests/ cleaning rules, leakage guard, CV folds, intervals, pipeline smoke
notebooks/ demo: train the model and query real routes
app/ Streamlit demo with estimate and 80% range
reports/ generated tables and plots
docs/ full methodology
The full write-up is in docs/methodology.md: cleaning rules with row counts, EDA findings, every CV experiment, the error analysis, the v2 changes and interval calibration, and the limits of these results.
Data: NYC TLC yellow taxi trip records (source). The data is downloaded by script and not redistributed here; check the TLC terms before reusing it. The code is MIT-licensed (LICENSE).