CBC Reference Model: Retail Demand Forecasting (NYC Taxi hourly)
Pre-trained reference model for the CBC MLOps 100-Day Track (Capstone 3). Published twin of ML Development Capstone 3.
Model details
- Type: XGBoost regressor (n_estimators=300, max_depth=4) on 12 past-only lag/rolling/calendar features of an hourly demand series. Seed 42, CPU, 0.46 MB.
- Framework: xgboost 3.2.0 · Serialization: joblib (full XGBRegressor;
.predict(DataFrame)-> predicted trips). - C3 is deliberately classical: a univariate LSTM (ML Dev Day 74) loses to this GBM on a medium-size, strongly-seasonal series.
Intended use
A next-hour demand estimate to support staffing/dispatch planning. Decision support, not an automated control signal. Teaching/reference artifact.
Training data
NYC Yellow Taxi trip records, January 2024, aggregated to a 744-hour count series (a stand-in for retail demand). Public NYC.gov open data, no PII. NYC.gov Terms of Use (not CC0).
Metrics (untouched holdout + walk-forward, evaluated once)
| Model | Holdout MAE | RMSE | MAPE | R2 |
|---|---|---|---|---|
| previous-hour naive (lag_1) | 639.67 | 831.09 | 31.24% | - |
| same-hour-last-week naive (lag_168) | 270.05 | 418.94 | 7.82% | - |
| XGBoost | 279.39 | 389.85 | 12.24% | 0.9716 |
On a single holdout the model wins RMSE but trails the strong same-hour-last-week naive on MAE. The honest headline is walk-forward (the reliable estimate): XGBoost MAE 384.20 vs that naive 427.32 (per-fold [600.3, 402.5, 399.1, 252.6, 266.5]) — the model beats it on the estimate that matters. Top features: lag_1 0.51, lag_168 0.25, hour 0.11, lag_24 0.04.
How to load and predict
import joblib, json, pandas as pd
from huggingface_hub import hf_hub_download
model = joblib.load(hf_hub_download("careerbytecode/mlops-ref-retail-demand", "model/pipeline.joblib"))
sample = json.load(open(hf_hub_download("careerbytecode/mlops-ref-retail-demand", "sample_input.json")))
trips = float(model.predict(pd.DataFrame([sample]))[0])
print(trips) # predicted next-hour demand
Input schema: 12 past-only features (lags 1/2/3/24/168, rolling mean/std, hour, day_of_week, is_weekend, day_of_month) computed from real history.
Limitations
- Hourly demand is strongly weekly-periodic — the one-line same-hour-last-week naive is a very strong bar; the model only modestly beats it (walk-forward).
- Trained on a single month (744 hours); longer seasonality (holidays, weather, trend) is not represented and will drift.
- Features are past-only; serving must supply the same 12 values from real history. Reference/teaching artifact only.
© 2015-2026 CareerByteCode. All rights reserved. | CC BY-NC-SA 4.0 (docs), MIT (code) | Authored by Raghavendra R, Platform Owner CareerByteCode, Solution Architect