Bike Sharing Demand Prediction using Machine Learning

Bike-sharing operators rebalance stations with trucks every day, and every empty station at rush hour is a lost customer. Demand follows strong, learnable structure — twin commute peaks on working days, a broad afternoon hump on weekends, summer highs and winter lows, and a roughly 50% collapse in rain. This project trains an XGBoost regressor on log-transformed hourly counts from the well-known UCI Capital Bikeshare dataset to learn those patterns, with RMSLE as the objective (the metric the field uses, since errors matter relative to demand level). The demo renders a 48-hour forecast under selectable weather and a per-hour predictor with driver breakdowns. Validation is strictly chronological — train on 2011, test on 2012 — because shuffling time series leaks the future. Built with standard tools (Python, XGBoost, scikit-learn).

1. The UCI bike-sharing dataset (17,389 hourly records, 2011–2012, with weather and calendar fields) is loaded and checked for gaps.
2. The rental count is log1p-transformed to stabilize variance across quiet nights and rush hours.
3. Data is split by year — train on 2011, validate/test on 2012 — the honest chronological protocol.
4. An XGBoost regressor is trained with RMSLE objective and early stopping on validation error.
5. The 2012 holdout is evaluated once: RMSLE 0.32, R² 0.91 are the design targets; errors are broken down by hour, season and weather.
6. The web demo loads the saved model for the 48-hour forecast view and the per-hour predictor.

Tech stack:

• Python 3, XGBoost
• scikit-learn (metrics, validation)
• NumPy, pandas, Matplotlib
• Jupyter Notebook (training & evaluation)
• HTML5 canvas + JavaScript (forecast demo)
• UCI bike-sharing dataset (Capital Bikeshare)

Dataset & model details

• Dataset: UCI "Bike Sharing Dataset" — 17,389 hourly records (731 days), Capital Bikeshare, Washington D.C., 2011–2012; features: hour, weekday, month, season, temperature, humidity, windspeed, weather situation, working-day/holiday flags.
• Task: Regression; input = calendar + weather features, output = hourly rental count.
• Model: XGBoost regressor on log1p(count) with RMSLE objective; early stopping on validation RMSLE.
• Metrics: RMSLE, RMSE, R² on the 2012 holdout — RMSLE 0.32 and R² 0.91 are the design targets for the built-to-order training run, stated honestly and never claimed as measured.

• [Hourly demand regressor] XGBoost on log1p(rental count) with RMSLE objective — errors penalized relative to demand, as operators care about.
• [48-hour forecast demo] Typical vs forecast demand curves with a weather selector showing how rain halves the curve.
• [Per-hour demand predictor] Enter hour, day type, temperature and weather → predicted rentals with driver breakdown.
• [Demand-pattern analyses] Working-day vs weekend hourly curves, monthly seasonality bars, temperature and weather response — all visualized.
• [Chronological validation] Train on 2011, test on 2012: the model is evaluated on a full unseen year, no leakage.
• [Feature importance study] Hour, temperature and working-day flag ranked by gain, matching the demand literature.
• [Log-scale diagnostics] Actual-vs-predicted scatter on the log scale with RMSLE bands, documented in the notebook.

• Data pipeline & exploratory demand-pattern analyses
• Complete training & evaluation Jupyter notebook
• Tuned XGBoost demand regressor with inference code
• Interactive web demo (48-h forecast + per-hour predictor)
• Project report PDF (demand theory, methodology, honest evaluation)
• PPT presentation for final review
• Viva Q&A preparation document (RMSLE vs RMSE, log transform, chronological splits, gradient boosting)