IA idea · Sport
Can a simple rating system predict football results?
Research question
How accurately does an Elo rating system, built from past seasons of [a league], predict match outcomes, and which value of the update constant K gives the best predictions?
Adapt it: change the place, the data or the comparison until the question is yours.
Why it makes a good exploration
Elo ratings rank chess players and are used for football and other sports. Building one yourself, then measuring how good its predictions are, is a genuine data-science project with a parameter to optimise.
The mathematics you'll need
- The logistic expected-score formula E = 1/(1 + 10^(−Δ/400))
- Iterative updating (a recurrence)
- Measuring prediction quality (e.g., mean squared error of probabilities)
- Optimising a parameter by comparison
Course labels show where a technique sits; using maths from outside your course is fine if you explain it clearly and say it is new to you.
Where the data comes from
Use several seasons of results from Football-Data.co.uk (which also include bookmakers' odds for comparison).
- Football-Data.co.uk — Match results, shots, cards and bookmaker odds for 25+ seasons of European leagues, CSV.
Cite every source in a footnote where you use it and in your bibliography. Check the licence of any dataset you download.
A possible outline
- Explain the Elo formula and its logic.
- Build the ratings in a spreadsheet.
- Predict each match before updating.
- Score predictions and optimise K.
- Compare with bookmakers' implied probabilities and reflect on draws and home advantage.
Pitfalls that cost marks
- Using results to predict matches that were used to build the ratings.
- Ignoring draws.
- Not explaining the logistic formula.
Showing personal engagement
- Predict next weekend's matches and check.
- Add a home-advantage term and justify its size.
- Compare your model's accuracy with your own guesses.
See Criterion C: personal engagement for what examiners look for.
Taking it further
Add a margin-of-victory adjustment and test whether it improves predictions.