Tardis
Cleaning SNCF train-delay data, training a model to predict delays, and a Streamlit dashboard. The model barely beat guessing, and the page says so.
- Period
- Apr 2025 – May 2025
- Role
- Data cleaning, model and dashboard
- Team
- 3 people
- Status
- archived
Stack and tags
Can we predict how late a French train will arrive? Tardis takes the SNCF data on train punctuality, cleans it, trains a model, and puts the result in a small web app where anyone can browse the statistics or try a prediction.
How it works
Most of the work is cleaning: broken lines, numbers stored as text, impossible dates, and station names with typos (Dinkerque for Dunkerque). Each repair is documented step by step. A model then learns from the clean data, and the dashboard shows delays by station and by hour, with a form to ask for a predicted delay.
The result is reported honestly: the model did not beat simply answering with the average delay. The notebook keeps that finding instead of hiding it.
What I took from it
- A model that fails to beat the average is still a finding: it says the inputs are not the story, or that the preparation needs another look.
- Cleaning data is most of the job, and it deserves to be written down as carefully as the model.
