Open-source codebase for short-term ridership prediction

Models Implemented

• ARIMA
• SARIMA
• Multilayer Perceptron (MLP)
• Convolutional Neural Network (CNN)
• Long-Short-Term memory (LSTM)

We kindly request that users of this open-source platform cite this paper in any resulting publications or works that make use of the platform:

Caicedo, J. D., Guirado, C., González, M. C., & Walker, J. L. (2025). Sharing, collaborating, and benchmarking to advance travel demand research: A demonstration of short-term ridership prediction. Transport Policy, 171, 531–541. https://doi.org/https://doi.org/10.1016/j.tranpol.2025.06.009

Overview

This project offers code and notebooks for forecasting transit ridership. It supports both static training and online updates as new data arrives.

Repository Layout

.
├── run.py                # main entry point
├── data.py               # data loading and feature engineering
├── models/               # ARIMA, SARIMA, Dense, CNN, LSTM classes
├── preprocessing/        # notebooks and helpers for preparing data
├── experiments/          # exploratory notebooks and trials
├── results_plots.py      # plotting utilities
├── settings.yaml         # default configuration
└── requirements.txt      # list of dependencies

utils.py provides helper functions for reading/writing YAML or JSON, and time_measure.py implements a simple TicToc timer.

Getting Started

The repository uses Python and only requires a few packages. You can set up the environment with pip:

pip install -r requirements.txt

The default configuration lives in settings.yaml. It specifies the input dataset location, aggregation level, model type and other parameters. You can run the code with the default settings by executing:

python run.py

Command line options allow you to override the configuration. Common flags are:

• --model – choose between dense, cnn, lstm, arima or sarima.
• --online – enable online (rolling) training.
• --stations – list of station names for single‑output mode.
• --aggregation – temporal aggregation such as hour or day.

Example:

python run.py --model cnn --aggregation day --stations "(1234) example station"

Predictions and timing information are saved under the output directory.

Next Steps

Explore the notebooks in experiments/ and preprocessing/ to see how the raw data was transformed and how different models were evaluated. Tweaking settings.yaml is an easy way to try new parameters or models.