Flowcast

An end-to-end AWS machine-learning platform that forecasts hourly yellow-taxi pickup demand for every NYC taxi zone 1, 6, and 24 hours ahead, and makes the model's performance inspectable by zone, horizon, and hour.

Live public site · Aug – Oct 2026

My role

Solo build: data pipeline, feature engineering, model training and evaluation, cloud serving path, and the web product.

Live website Model card Performance dashboard GitHub showcase

The New York City skyline and Brooklyn Bridge at night.

NYC taxi zones forecast: 262Every canonical five-borough TLC zone.
Lower held-out WAPE: 24.8%0.186 vs. 0.247 for a 168-hour seasonal-naive baseline, Mar–Jul 2026 test period.
Training examples: 13.8MZone × hour × horizon examples, built from 116.3M raw trip records.

The problem

Monthly trip files tell you what already happened. They do not answer the operational question: how many pickups should a zone expect in the next hour, six hours, or day, and how much should you trust that number?

A forecast is only useful if it beats a sensible baseline, respects time when it is evaluated, and shows its uncertainty. I wanted a system where every one of those claims could be checked from the product itself.

My contribution

I built Flowcast alone, from raw data to the public interface: the S3 data lake, leakage-safe feature engineering, the SageMaker training and evaluation pipeline, the serverless inference path, and the Next.js product with its map, dashboard, and model card.

What I built

Flowcast joins 31 months of NYC TLC yellow-taxi trips (January 2024 through July 2026, 116.3M trips) with NOAA hourly weather, then trains one global XGBoost model across all 262 zones and all three horizons. A quality gate only registers a model that beats the baseline by a set margin.

The web product connects the forecast to evidence: a citywide map for any origin hour and horizon, per-zone drivers from SHAP, calibrated prediction intervals, a held-out performance dashboard, and a public model card.

Citywide MapLibre forecast map with historical replay against actuals and the baseline.
Per-zone drivers from SHAP and prediction intervals calibrated on validation data only.
Performance dashboard with error by horizon and borough, residuals, and the largest misses.
Public model card covering data, features, time-based splits, the baseline, the quality gate, and limitations.

How it works

Each stage hands a versioned artifact to the next, and the model never reaches the endpoint without passing its evaluation gate.

TLC trips + NOAA weather
S3 data lake
SageMaker Pipeline: features → XGBoost → evaluation
Quality gate + Model Registry
Serverless endpoint → Lambda → API Gateway
Next.js + MapLibre product

Data

Monthly TLC trip records and NOAA GHCNh hourly weather in a private S3 bucket, aggregated to zone-hour demand.

Features

Shifted lags and rolling windows, train-only encodings, and calendar and weather features, built so no future information leaks into a prediction.

Training

One global XGBoost model in a SageMaker Pipeline with chronological train, validation, and test periods.

Evaluation

Test metrics against a 168-hour seasonal-naive baseline, validation-calibrated intervals, SHAP, drift checks, and a gate requiring at least a 5% improvement.

Serving

A serverless endpoint behind Lambda and API Gateway that scores all 262 zones in one batched request.

Flowcast performance dashboard showing actual and predicted pickups, WAPE, MAE, RMSE, and the improvement over the baseline for the selected window. — The public performance dashboard: held-out results against the seasonal baseline, filterable by window, horizon, borough, and zone.

Key decisions

Compare against a baseline people already use.

The benchmark is the same zone and hour one week earlier. It is simple, hard to beat in a city with strong weekly rhythms, and makes the improvement easy to interpret.

Split by time, never at random.

Training, validation, and test periods are chronological, and intervals are calibrated on validation only, so the reported test results reflect genuinely unseen months.

One global model instead of 262 small ones.

A single XGBoost model learns shared patterns across zones and horizons, which keeps sparse outer-borough zones from being modeled on a handful of trips.

Make every claim inspectable.

Predictions, actuals, baseline values, intervals, and the model version stay visible in the product, so a reader can check the results instead of trusting a headline number.

Outcomes & evidence

On the held-out test period (March 1 through July 31, 2026), Flowcast reached a WAPE of 0.186 against 0.247 for the seasonal-naive baseline, a 24.8% relative reduction. It beat the baseline at every horizon (1h, 6h, and 24h) and in every borough, and the 80% and 95% prediction intervals covered 83% and 96% of test outcomes.

The 13,779,890 training examples are zone-hour-horizon rows, not trips; the raw data behind them is 116.3M trip records. Those numbers come from the project's own model card and evaluation artifacts.

Scope & limits

These are offline, project-reported results, not measured operational impact; no taxi operator uses Flowcast. TLC data is published with a delay and covers yellow taxis only, so this is not a real-time citywide mobility feed.

Historical backtests use the weather that was actually observed for the target hour. An operational system would need forecast weather available at prediction time, which the model card states plainly.