Urban Bike-Sharing: A Spatio-Temporal Analysis of Washington, DC

Overview

This project analyzes Washington, DC’s bike-sharing system using spatio-temporal models to identify demand patterns and optimize urban mobility. By integrating bike rental data with weather and temporal variables, it evaluates the impact of external factors on usage and compares predictive models. Developed as part of the course Statistics for High Dimensional Data and CompstatLab (Master’s in Computer Engineering).

Objectives

• Investigate how weather and time variables influence bike-sharing demand.

• Compare Dynamic Coregionalization Model (DCM) and Hidden Dynamic Geostatistical Model (HDGM) for predictive accuracy.

• Propose data-driven strategies to optimize bike allocation and service efficiency.

• The key questions are:

  • How do weather and events (like holidays) affect bike usage?

  • Which spatio-temporal model (DCM or HDGM) predicts bike usage more accurately?

Methodology

• Data Preprocessing & EDA: Cleaned and analyzed data using Python (Pandas, NumPy, Matplotlib/Seaborn).

• Modeling:

  • DCM:
    • Hierarchical model capturing spatial-temporal dependencies.
    • Accounts for both spatial and temporal correlations.
    • Uses two latent components: one for spatial dependencies w(s,t) and one for temporal dynamics z(t).
    • More complex but flexible.
  • HDGM:
    • Latent process model for dynamic geospatial patterns.
    • Simpler, with only one latent component z(s,t) for spatio-temporal dynamics.
    • Less prone to overfitting.

• Tools: Implemented in MATLAB (D-STEM v2) for modeling, validated via cross-validation.

Data Exploration

Hourly Rentals

Figure 1: Distribution of rentals across hours of the day.

Daily Rentals

Figure 2: Distribution of rentals across hours of the day.

Monthly Trends

Figure 3: Monthly usage variations across 2023.

Seasonal Trends

Figure 4: Comparative demand across seasons.

Key Findings

• Temperature, precipitation, and UV index are key demand predictors.
• Longer trips on weekends/holidays reflect recreational vs. commuting use.

Future Work

Expand with socio-demographic data, real-time traffic integration, and multi-year analysis.

References

• Finazzi, F., Wang, Y., & Fassò, A. (2021). D-STEM v2. Journal of Statistical Software, 99(10).
• Capital Bikeshare Data | Visual Crossing Weather