Tennis match prediction

The problem

I wanted to know which stats actually decide a tennis match, and whether a model could call the winner before the first serve. I had written code in classes but had never built my own model.

What I did

• Pulled more than 73,000 ATP matches going back to 2000 out of large match databases with SQL.
• Built features from serve stats, break-point conversion, surface history and recent form, then trained an XGBoost model on them.
• Tested it on whole seasons the model had never seen, not a random split, because that is how it would be used.
• Filled the gaps with Coursera, YouTube, research papers on sports prediction, and a lot of watching real matches.

Result

Version one scored over 90%. It was cheating: some features were stats that only exist after a match ends.

Rebuilt on pre-match information only, the model is right 68% of the time on held-out seasons. Most of that comes from matches with a clear favorite, so the real work now is the close, 50/50 matches.

What I learned

• A great accuracy number is a reason to check your data, not to celebrate.
• Always compare against a simple baseline, like picking the higher-ranked player.
• Validate a model the way it will actually be used.