Winning project in the University of Manchester MathSoc x AutoTrader Hackathon, March 2024
Our task was to predict how long it would take a car listed on AutoTrader's platform to sell. With a tight deadline of 10 days, and 250,000 rows of data, we aimed to build a model to answer this question quickly and accurately on unseen cars.
Who are we?
- Ioan Gwenter (www.linkedin.com/in/ioan-gwenter, www.github.com/ioan-gwenter)
- Lourenço Silva (www.linkedin.com/in/lourencofsilva, www.github.com/lourencofsilva)
- Tom Cassar (linkedin.com/in/tom-cassar, www.github.com/tcassar)
We studied the machine learning pipeline, and immediately built a model that worked brilliantly - however, we had an issue with data leakage. So, we restarted and did some in-depth data exploration (see ./exploration).
We experimented with decision trees, random forests, and artificial neural networks ./models. We looked at published papers to see if regression or classification would be a better approach. We tried automated ML platforms such as Edge Impulse, and FeatureTools - an automated feature engineering platform. We looked at various forms of hyper-parameter optimisation and model evaluation.
Selected to showcase our work, we presented our findings to AutoTrader and fellow finalists. The presentation, complete with slides and speaker notes, is available in ./slides.
It was extremely interesting to see the approaches that the other teams had taken. Everyone had a different approach to feature engineering that came with their own positives and negatives.