?
This project was created out of my love for MMA; the necessity to do something to stay awake between fights as a European MMA fan and the hope to finally win with betting.
Fight Forecast leverages the power of machine learning to analyze historical UFC fight data,
turning it into a predictive tool for upcoming events. These models are meant to offer
insightful predictions and to be used as a companion for any MMA bettor seeking a data-driven edge.
This is all for fun, this model doesn't guarantee anything, I've made this model as a fun side project in my free time and I hope it is of use to some people.
Betting Columns
Kelly: Specifies the betting amount on the predicted winner according to 1/10th of the Kelly criterion. If no value is shown then the predicted winner isn't considered a viable bet according to the Kelly criterion
Heuristic: Bets according to the Kelly model with the exception to fights that fall under certain rules (e.g. at least 1 of the fighters only has 1-2 UFC fights). This is most likely not an optimal betting strategy. It probably doesn't lose you money, but it doesn't make you a lot of money either.
Model details
Basic model 76-77% test, 77-80% train (Used to be ~69% up until Max vs Zombie.).
Model is created using XGBoost.
Data used is everything from UFCstats.com, Historical betting odds & Location elevation.
Number of core features is around 80, total number of features after engineering is around ~3000.
IMPORTANT: This model does not take into account the following important and obvious features: Short notice fight, Home advantage, Missing weight, Injuries or anything else health related. (e.g. Weidman breaking his leg, my prediction didnt reflect this at all...
The forecast will update every monday and then again around friday at the very least.
Acknowledgments
I Want to thank Dan McInerney (danhmcinerney@gmail.com) from mma-ai.net for continuous exchange of information and help with reducing overfitting.