CycleDemand: Bike Sharing Demand Forecasting

Objective

The goal of this project is to forecast daily bike rental demand based on historical usage patterns and weather conditions.
Accurate forecasts support operational planning (e.g., bike availability, staffing).

---

Dataset

• Source: UCI Bike Sharing Dataset (2011–2012)
• Granularity: Daily rentals (y)
• Features:
  • Calendar: year, month, day-of-week, holiday, working day, season
  • Weather: temperature, humidity, windspeed
  • Lagged demand: y_lag1, y_lag7, rolling averages

---

Methodology

1. Exploratory Data Analysis (EDA)

   • Visualized seasonal trends, weekday vs weekend, and weather impacts.

2. Baselines

   • Naïve (yesterday’s rentals)
   • Seasonal naïve (7-day lag)

3. Machine Learning Models

   • LightGBM
   • XGBoost

4. Evaluation Metrics

   • MAE (Mean Absolute Error)
   • RMSE (Root Mean Squared Error)
   • WAPE (Weighted Absolute Percentage Error)

---

Results

Metrics Comparison

Model	MAE	RMSE	WAPE
Naive (y_t-1) [VALID]	728.4	1032.9	0.1053
Seasonal Naive (y_t-7) [VALID]	802.0	1115.2	0.1159
Naive (y_t-1) [TEST]	951.7	1380.2	0.1843
Seasonal Naive (y_t-7) [TEST]	1457.6	2047.2	0.2822
LightGBM [VALID]	621.9	780.4	0.0899
LightGBM [TEST]	915.8	1145.8	0.1773
XGBoost [VALID]	792.3	930.9	0.1145
XGBoost [TEST]	1011.1	1197.7	0.1957

---

🧾 Explainability

SHAP Feature Importances

SHAP Beeswarm

---

💡 Key Insights

• Lag features dominate: y_lag1, y_lag7, and rolling averages are the strongest predictors.
• Weather effects matter: rentals rise with mild temperatures, drop with high humidity and extreme cold.
• Seasonality is clear: demand peaks in summer, weekends differ from weekdays.
• Model comparison:
  • LightGBM: best performance (RMSE ~1146, MAE ~916 on test).
  • XGBoost: competitive but slightly worse (RMSE ~1198, MAE ~1011).
  • Both beat naïve baselines by a wide margin.
• Explainability: SHAP confirmed historical demand and weather as top drivers of predictions.

---

Conclusion

• The project achieved forecast accuracy of ~900 rentals/day MAE.
• LightGBM provided the best balance of accuracy and interpretability.
• The pipeline demonstrates EDA → baselines → models → explainability in a reproducible workflow.
• This level of forecasting can support fleet allocation, staffing, and planning for bike sharing services.

---

Next Steps

• Tune hyperparameters for further improvement.
• Explore hourly demand forecasting.
• Package the model into a deployable API or streaming service.