Speed vs. Sense: TfL Cycle Hire Route-Choice Analysis

This repository contains the code and notebooks used for the project/paper “Speed vs. Sense: The Hidden Logic of Cyclist Routes” (1c_paper_andrii_kolisnyk.pdf).

At a high level, it:

• Takes TfL Cycle Hire trip records (two 2-week windows: winter vs. summer).
• Converts docking-station identifiers to lat/long coordinates using the live station feed.
• Enriches trips with TfL Journey Planner API estimates for:
  • theoretical cycling time (fastest route)
  • public-transport time (alternative)
• Builds a logistic regression model to explore which contextual factors are associated with choosing the fastest vs. slower / “scenic/ideal” route.

What the repo is (and isn’t)

• This is primarily an exploratory, reproducible research pipeline.
• “Fastest route vs. scenic route” is an operational definition derived from time deltas: trips within +10% of the API’s theoretical cycling duration are treated as “fastest-route” choices.
• Because the TfL Journey Planner API does not support historical queries beyond a short window, the notebook maps each historical trip date to the nearest upcoming date with the same weekday before calling the API.

Data sources

• TfL Cycling Data Portal: trip extracts (stored under data/raw)
  • January (winter): data/raw/cycling_jan15-31.csv
  • July (summer): data/raw/cycling_jul15-31.csv
• Cycle hire station feed (XML): data/raw/livecyclehireupdates.xml
• TfL Journey Planner API (used during enrichment; see utils/tfl_api.py)

Repository layout

tfl-analysis/
	1c_paper_andrii_kolisnyk.pdf
	notebooks/
		1_bike_data_extraction.ipynb
		2_cyclists_efficiency_modeling.ipynb
	utils/
		tfl_api.py
	data/
		raw/                         # input CSVs + station XML
		processed/
			cleaned_data_sample10.csv  # stratified 10% sample (from notebook 1)
			api_processed/
				sample_10.parquet        # API-enriched dataset (from utils/tfl_api.py)
		result/                      # optional exports (see notebook 2)
	logs/
		tfl_api.log                  # API call logs (when running utils/tfl_api.py)

Quick start

1) Create env + install dependencies

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

2) Build the cleaned stratified sample (Notebook 1)

Open and run: notebooks/1_bike_data_extraction.ipynb

This notebook:

• loads the two raw CSV extracts
• removes trips with the same start/end station
• creates API-compatible date/time fields
• joins station coordinates from the XML feed
• writes a stratified sample to: data/processed/cleaned_data_sample10.csv

3) Enrich the sample using TfL Journey Planner API

Run the API script:

python utils/tfl_api.py

It reads data/processed/cleaned_data_sample10.csv and writes:

• data/processed/api_processed/sample_10.parquet

Notes:

• The API has rate limits (the code includes sleeps/retries and logs to logs/tfl_api.log).
• The script currently embeds an API key; if you fork this, you should replace it with your own key.

4) Fit the logistic regression + generate diagnostics (Notebook 2)

Open and run: notebooks/2_cyclists_efficiency_modeling.ipynb

This notebook:

• loads the API-enriched parquet
• engineers variables used in the paper (time-of-day class, distance class, deltas)
• fits a logistic regression with treatment-coded categorical variables
• produces odds ratios + residual plots
• includes optional exports to data/result

Model definition (as implemented)

• Dependent variable: cyclist_class (binary)
  • 1 if (real_duration - cycling_api_duration) / real_duration < 0.1
  • else 0
• Key predictor: cycling_public_delta_pct (cycling vs public transport)
• Controls (categorical): type_of_day, time_class, season, distance_class, bike_model

Paper

For methodology rationale, assumptions, limitations, and findings, see:

• 1c_paper_andrii_kolisnyk.pdf

Support

If this repository or paper helped you, please consider starring the repo. It helps others find it and supports continued work.

Acknowledgements

Powered by TfL Open Data.