Round Robin Tournament Forecasts for Tour Finals

[mcekovic]

• Predict all outstanding matches in groups, calculate group position for each group match outcome combinations and base probabilities to qualify to first KO round.

[swissknife007]

Will love to go over your code for the match result prediction algorithm and contribute to it perhaps !

[mcekovic]

It starts from https://github.com/mcekovic/tennis-crystal-ball/blob/master/tennis-stats/src/main/java/org/strangeforest/tcb/stats/service/MatchPredictionService.java

[mcekovic]

It yields pretty decent results, much better then some commercial tennis prediction services that are available. When comparing the prediction to the bookmaker odds, you can see that they match pretty well. It can predict match outcome with ~70% accuracy (~68% for best-of-3 and ~75% for best-of-5). There are areas for improvement, but it already is very near to the theoretical limits, due to the natural randomness of the tennis match outcomes (maybe can be improved by at most couple of percents). Areas of improvement I see:

• Use pattern detection neurons/prediction items (examples: physical fatigue after long matches, psychological fatigue after winning over favorites, etc...).
• Use per-player or per-common-player-style training/tuning (like Big Server, Aggressive Baseliner, Defensive Baseliner, Serve & Volleyeler...)
• Use more complex neural network layers (instead of current relatively simple 3-layers: 1 input, 1-hidden, 1-output)
• Add Qualifying/Challenger/Futures matches into Elo calculation, for better Elo ratings of Challenger-level players that occasionally enter ATP tournaments
• Add cross-combination of serve and return Elo ratings as one prediction item/neuron (preliminary results say it will be useless and will not be picked by training/tuning as relevant - so low priority)