CityBike Lisbon: Location Data Modelling with Python

Project

This project explores the influence of surrounding businesses on bike availability at Lisbon CityBike stations. Using data from the CityBike and Foursquare APIs, I performed multivariate regression analysis to assess how proximity to bars, cafes, and other venues impacts station usage.

Goals

• Identify predictor variables that influence bike availability at Lisbon CityBike stations.
• Visualise trends between station-level bike availability and surrounding locations
• Use multivariate regression analysis to determine whether location context meaningfully predicts bike availability.

Data

Tech Stack

• Python (pandas, seaborn, matplotlib, statsmodels)
• APIs: CityBike, Foursquare
• PostgreSQL
• Jupyter Notebooks

Process

Step 1

• city_bikes.ipynb

• Acquired CityBike Data from CityBike API

  • Resulting 195 bike station coordinates

• Parsing JSON File data into Dataframes for cleaning

• CityBike API Data Timestamp = 12:39 AM Sunday, July 27 2025

Step 2

• yelp_foursquare_EDA.ipynb

• Acquired location data from FOURSQAURE APIs within 1,000metres of Bike Stations

  • Data parameters included;

    • Location Name
    • Distance From CityBike Station
    • Location Category
    • Rating
    • Popularity
    • Hours
    • Hours Popular
    • Price

  • 9,662 Locations within 1,000 metres radii of Bike Stations

    • Grouped by Bike Station, visualisation methods performed for cleaning, aggregation and modelling

• Issue Unable to access YELP API without Business Contact

  • YELP API Step Ignored

Final cleaned dataset:

• 176 Bike Stations
• 9,662 surrounding locations (filtered)

SQL Database

• Created SQL Database for future access
  • Created, linked and stored data in normalised tables

Step 3

• Exploratory Data Analysis

  • Conducted visual inspections with seaborn/matplotlib pairplots

    • Observed null or weak visual relationships

  • Generated Pearson R correlations for proposed linear relationships & variance inflation factor values for Multicollinearity

  • Observed weak linear correlations between x-variables and y-variables

    • y-variables
      • free_bikes, total_slots

  • Cleaning Procedure

    • Foursquare API Limitation:
      • API limit at 50 locations per bike station, which introduces selection bias.
    • Filtered Locations distance < 298 metres to remove selection bias of 50 location limitations
      • Bike Station Maximum locations = 49
    • Cleaned Quantity : Bike Stations = 176

Step 4

• Statistical Modelling

  • Mulivariate Regression Models

    • Dependent Variables
      • Free Bikes
      • Total Bike Slots
    • Full model with all business categories
    • Refined model using only top independent variables
      • bar_count
      • cafe_counts
    • Adjusted R² improved in simplified model

  • Linear Regression Models

    • Independent Variable
      • Average Popularity of locations
    • Dependent Variables
      • Free Bikes
      • Total Bike Slots
    • Strong correlations
      • Statistically Insignificant
      • Maximum of 1.4% explanatory power on the variability of dependent variables

Results

API Quality

• Foursquare’s API returned a maximum of 50 location results per coordinate query, which introduced selection bias in areas with dense points of interest.
• To mitigate this, we applied a filter of < 298 metres, reducing the max results per Bike Station to 49.
• Final cleaned dataset includes 176 Bike Stations with confirmed nearby locations under this threshold.

Visual Observations

• Pairplots showed no clear visual relationship between business categories, average popularity or average ratings, and the availability of Free Bikes or Total Bike Slots.

• This held true both before and after bias correction, suggesting a weak influence of surrounding locations on bike availability.

• Cleaned Dataframe Scatterplot Visualisations
  

  • Scatterplot of Free Vikes Vs Ratings

  

  • Scatterplot of Free Bikes Vs Categories

  

  • Pairplot of all Bike Stations

OLS Regression Model Findings

• Pearson R correlation and VIF analyses confirmed low multicollinearity for business categories, with a proposed threshold of 5. However, little predictive strength in further analysis.

  • VIF analysis did find high multicollinearity among ratings and popularity. These variables removed from Multivariate Regression Analysis.

• A linear regression model was applied using Average Popularity of Locations as the independent variable, and Free Bikes and Total Bike Slots as the dependent variables.

  • Weak linear relationships were observed between top independent variables and dependent variables.

• Selecting Top Independeent Variables:

  • Coefficients
    • Each additional bar within 298 metres is associated with 1.67 more free bikes.
    • Each additional café is associated with 0.74 fewer free bikes.
  • p-value
    • Both Correlations Statistically Significant at the level of 5% alpha
  • Adjusted R-squared of multivariate models increased:
    • The model explains approximately 17.9% of the variance in free bike availability, indicating low predictive power.

Conclusion

• The Foursquare API provided geospatial location context, but within Lisbon city, the nearby venue data utilised does not show strong predictive influence on city bike station metrics.

• No strong predictive relationships or correlations were observed between Location Categories or Location Ratings with Total Bike Slots or Free Bikes of CityBike Stations.

  • Location Categories are a poor predictor of both Bike dependent variables
  • Location Ratings & Popularity are a poor predictor of both Bike dependent variables

• However, bar_count and cafe_counts offered the strongest available predictive value for Free Bikes accounting for approximately 17.9% of the variance.

• Alternative predictor variables and features not present in this analysis may be more valuable in predicting bike availability than nearby venue types or ratings

Future Goals

• Additional Independent Variables

  • Examples
    • Time Of Day
    • Season
  • CityBike API Data Timestamp = 12:39 AM Sunday, July 27 2025
    • Further Goal to observe trends across daily, weekly & monthly timestamps

• Cleaning further of categories

  • Top 4 cleaned categories utilised from CityBike API
    • Restaurant
    • Bar
    • Cafe
    • Coffee
  • Further cleaning and grouping of broader categories
    • Examples
      • Public / Government
      • Retail
      • Hospitality

• Additional Statistical Analysis

  • Statistical Testing
    • Models
      • Plot residuals vs fitted values
      • Generate Q-Q plot of residuals
      • Add a line of best fit to scatterplots using sns.regplot()
    • T-Tests to identify acquired sample data of locations as a reflection of true populations.
      • As a result of restricted FOURSQUARE API Response limits.
    • However, given poor predictor variables, significance of successful Statistical Testing minimal due to infered results

Challenges

• Selection Bias in FOURSQUARE API Reponses.

  • Restricted responses from FOURSQUARE API required a limitation of 50 locations for Bike Station Coordiantes.
  • Restricted data available to assess full population of locations surrounding Bike Station, requiring cleaning and reduction of location distance threshold.

• Low Predictive Strength of Independent Variables.

  • Low/Weak linear relationships from all geospatially available data with our CityBike Stations dependent variables.
  • Resulting conclusions are low model effectiveness and future goals to identify alternative independent variables.

• Multicollinearity

  • Strong multicollinearity present among top linear correlated variables, Average Rating and Average Popularity.
  • Separation of multicollinearity variables was required to meet assumptions for accurate model predictions.