Hourly Bike Rentals Forecasting in Florida

Submission Date

July 12, 2025

By Carllos Watts-Nogueira

As the sole developer and analyst, I designed and implemented an end-to-end machine learning pipeline in the context of my academic journey at the University of San Diego’s AI and Machine Learning program, powered by Fullstack Academy (Anticipated Graduation Date: October 2025).

The project exemplifies the hands-on skill development central to the bootcamp, covering topics such as applied data science, machine learning, deep learning, NLP, and generative AI. I engineered features, cleaned and scaled data, built and tuned regression models, conducted exploratory testing with PCA and polynomial transformations, and evaluated performance using industry-standard metrics.

Project Objective

Forecast hourly bike rental demand based on weather and time-based features, applying various regression techniques and modeling workflows.

Technologies & Skills Used

• pandas – data wrangling, preprocessing
• scikit-learn – regression models (Linear, Ridge, Lasso, ElasticNet), pipeline building, scaling, encoding, PolynomialFeatures, PCA, GridSearchCV, cross-validation
• joblib – saving serialized models
• matplotlib – data visualization
• Feature Engineering – including:
  • Interactions like Temp × Humidity, Hour × Weekday, Wind × Visibility
  • Polynomial feature expansion (degree=2)
  • PCA dimensionality reduction
• Experimental Design – iterative testing of four pipelines, isolating the impact of PCA and engineered features
• Performance Evaluation – metrics: MAE, MSE, R²

Pipeline Versions

Version	PCA	Feature Interactions	Best Model	R²
01_base_model	❌	❌	Lasso	0.481
02_scaler_then_pca	✅	❌	Poly_Lasso	0.617
03_pca_only	✅	❌	Poly_Lasso	0.346
04_no_pca_with_interactions	❌	✅	Poly_Lasso	0.739 ✅

Highlight

The best-performing model was Poly_Lasso (R² = 0.739) using hand-crafted feature interactions without PCA, demonstrating the power of meaningful feature design over dimensionality reduction in this context.

Business Impact

These forecasts can support:

• Resource planning during peak hours
• Weather-based fleet allocation
• Staffing, dynamic pricing, and maintenance scheduling
• Integration into real-time decision systems

Next Steps

• Extend feature engineering (e.g., time-of-day bins, lag features)
• Try ensemble models (RandomForest, XGBoost)
• Create visual dashboards for ongoing monitoring

---

Additional Testing – Final Model Evaluation (August 2025)

A new round of testing was conducted to validate model performance using updated metrics and a refined pipeline.

📁 Folder: update_model_test_lasso

This folder contains:

• A new notebook with updated evaluation and cross-validation
• A saved production-ready model (best_model_lasso.pkl)
• A dedicated README summarizing the results and rationale

The best model from this round was Lasso Regression, selected for its balanced performance across MAE, MSE, and R², and its suitability for deployment.

Key Improvements

• Removed PCA from final pipeline
  This version uses only original features and engineered interactions.

• Enhanced Feature Engineering

  • Added temporal features: Weekday, IsWeekend, Month, Hour
  • Created meaningful interactions:
    • Temperature × Humidity
    • Hour × Weekday
    • Wind × Visibility
    • Temperature × Month
  • Scaled all numerical features using StandardScaler

• Polynomial Feature Expansion

  • Applied PolynomialFeatures(degree=3) to selected variables
  • Captured complex nonlinear relationships without overfitting

• Model Selection & Evaluation

  • Compared Linear, Ridge, Lasso, ElasticNet, and Polynomial Regression
  • Used GridSearchCV for hyperparameter tuning
  • Applied 5-fold cross-validation and test set evaluation
  • Final model: Lasso Regression with R² = 0.556 and lowest MAE

Business Impact

These forecasts support:

• Resource planning during peak hours
• Weather-based fleet allocation
• Staffing, dynamic pricing, and maintenance scheduling
• Integration into real-time decision systems