Seoul Bike Regression

Project Overview

This project analyzes Seoul bike-sharing demand with a regression-based workflow. The goal is to understand how rental bike demand changes with temperature, humidity, wind speed, visibility, season, holidays, working days, and other contextual variables.

The project focuses on interpretable statistical modeling rather than black-box prediction. It demonstrates the full modeling process from exploratory analysis to model specification, variable selection, diagnostics, and interpretation.

Dataset and Variables

The dataset contains hourly bike rental counts in Seoul together with weather and calendar variables. These variables are useful for studying mobility patterns and estimating demand under different environmental and temporal conditions.

Analysis Objectives

• Explore the relationship between rental bike count and weather/calendar features.
• Build interpretable linear regression models for demand analysis.
• Use model selection methods such as AIC to identify relevant predictors.
• Check regression assumptions using residual plots and diagnostic measures.
• Evaluate multicollinearity using VIF analysis.
• Summarize the main drivers of bike-sharing demand in a clear statistical report.

Methods

The project applies classical statistical modeling methods. These methods are useful because they provide interpretable coefficients and allow the analyst to understand how each variable contributes to the expected rental count.

Workflow

1. Loaded and inspected the Seoul bike-sharing dataset.
2. Cleaned variable names and prepared numerical and categorical features.
3. Explored demand patterns by weather, season, holiday, and functioning day.
4. Fitted baseline and extended linear regression models.
5. Applied AIC-based variable selection to refine the model.
6. Analyzed residuals and diagnostic plots to evaluate model assumptions.
7. Checked multicollinearity using variance inflation factors.
8. Interpreted the final model in terms of demand-driving variables.

Main Findings

The analysis showed that bike demand is strongly related to temperature, time-related variables, seasonality, and operational conditions. Demand generally increases under favorable weather conditions and changes substantially across seasons and hours.

The model also highlighted the importance of checking assumptions. Some patterns in the residuals indicate that bike demand is not perfectly linear and may require transformations, interaction terms, or more flexible models in a more advanced extension.

Diagnostics and Model Quality

Regression diagnostics were used to assess whether the selected model provides a reasonable statistical explanation. Residual plots, fitted-value comparisons, and multicollinearity checks helped identify strengths and limitations of the linear model.

• Residual diagnostics were used to inspect deviations from model assumptions.
• VIF analysis helped detect correlated predictors.
• AIC-based selection supported a more compact model specification.
• The final model is interpretable but not necessarily optimal for pure prediction.

Limitations and Possible Extensions

• A linear model may not fully capture nonlinear demand patterns.
• Interactions between temperature, season, and hour could improve the model.
• More advanced models such as random forests, gradient boosting, or time-series models could improve predictive performance.
• The project is mainly a statistical modeling and interpretation exercise, not a production forecasting system.

Outcome

This project strengthened my ability to perform a complete regression analysis workflow, including exploratory analysis, model building, model selection, residual diagnostics, and statistical interpretation.

It is a useful portfolio project for demonstrating classical statistical modeling, analytical reasoning, and clear communication of data-driven findings.