MODEL.details
Performance metrics on the last year of fights (data not used in model training)
Vegas Odds Performance
| Accuracy | 67.35% |
| Precision | 0.679 |
| Recall | 0.764 |
| F1 Score | 0.719 |
| Log Loss | 0.609 |
| Brier Score | 0.210 |
MMA-AI.net Performance
| Accuracy | 69.90% |
| Precision | 0.702 |
| Recall | 0.853 |
| F1 Score | 0.770 |
| Log Loss | 0.604 |
| Brier Score | 0.208 |
| Feature | Importance |
|---|---|
| Age Difference | |
| Submission Defense Opponent Average Difference | |
| Submission Landed per Minute Opponent Decayed Average Difference | |
| Age Ratio Difference | |
| Control per Minute Opponent Decayed Average Difference | |
| Head Accuracy Opponent Decayed Average Difference | |
| Reach Ratio Average Difference | |
| KO per Minute Opponent Decayed Average Difference | |
| Distance Landing Ratio Decayed Adjusted Performance Average Difference | |
| Significant Strike Accuracy Opponent Decayed Average Difference | |
| Significant Strike Landing Ratio Opponent Decayed Average Difference | |
| Head Landing Ratio Decayed Adjusted Performance Average Difference | |
| Significant Strike Landing Ratio Decayed Adjusted Performance Average Difference | |
| Clinch Landed Round 1 Adjusted Performance Decayed Average Difference | |
| Head Landing Ratio Opponent Decayed Average Difference | |
| Age Average Difference | |
| Days Since Last Fight Average Difference | |
| Takedown Accuracy Decayed Adjusted Performance Average Decayed Slope Difference | |
| Win Opponent Average Difference | |
| Takedown Landed per Minute Opponent Decayed Average Difference | |
| Distance Accuracy Opponent Decayed Average Difference | |
| Ground Landed Opponent Decayed Average Difference | |
| Control per Minute Decayed Adjusted Performance Average Difference | |
| KO per Minute Decayed Adjusted Performance Average Difference | |
| Reach Difference | |
| Submission Landed per Minute Decayed Adjusted Performance Average Difference | |
| Control Opponent Average Difference | |
| Ground Landing Ratio Decayed Adjusted Performance Average Difference | |
| Takedown Attempts Round 1 Adjusted Performance Average Difference | |
| Takedown Landing Ratio Decayed Adjusted Performance Average Decayed Slope Difference | |
| Days Since Last Fight Difference | |
| Leg Landed Round 1 Adjusted Performance Decayed Average Difference | |
| Takedown Landing Ratio Decayed Adjusted Performance Average Difference |
This curve shows how well the model's predicted probabilities align with actual outcomes. The closer to the diagonal line, the better calibrated the predictions.
Frequently Asked Questions
When do you post predictions?
Usually the day before or the day of the event.
How long have you been doing this?
About 4 years. Thousands and thousands of hours tuning, testing, and engineering. At this point I can absolutely chat your ear off about the intricacies of machine learning for sports prediction.
What is your background?
I was a hacker for about 15 years, now I'm an AI security researcher and engineer.
Where do you get the data?
Downloaded from various MMA stat sites. We start with 22 base stats like Strikes Landed and engineer those into 20,000 stats. Then we cut those down to about 10-20 of the most influential, predictive stats based on both manual expertise and statistical analysis.
What is the betting strategy you use?
Due to analysis by Chris from Wolftickets.ai, it was discovered that the most profitable strategy for machine learning sports prediction models is to include the odds in the dataset so the model will achieve accuracy better than the bookies then use 2 to 3 leg parlays based on the predictions.
Why is this free?
It may not be forever. This has been a 3 year endeavor to learn machine learning engineering and at this point I believe I am in possession of a world-class process and model.
How do you pick the parlays?
I scrape fight stats, engineer a ton of features, cut those features down to the 10-20 that are the best, then use Autogluon in a chronological 80/10/10 train/val/test split so the last 10% of data is never seen by the model and all model evaluations are reflective of how well it did on that last 10% of data.
What algorithm do you use?
I use an ensemble of traditional machine learning algorithms. Gradient boosted algorithms, neural networks, tree-based algorithms, and others. This is not an LLM tool, there is no ChatGPT here although I am exploring that as well.
What was the hardest part of this project?
Feature engineering but especially feature selection. This is what separates MMA-AI.net from all other MMA prediction tools (besides wolftickets.ai who helped me with this part). First, more features =/= better results. Second, feature selection is an art that is dependant on domain expertise and ENDLESS testing.