Model AUC
--
5-fold cross-validated
All-time RTP
--
96% = break-even threshold
Win Rate
--
--
Net Units
--
All-time P&L
Methodology
How It Works
01 / STYLE ENGINE
Predicting Fighting Style In-Fight
The model does not just classify a fighter's career style. It uses rolling lag features and shift-detection
to predict which style a fighter is likely to adopt entering a specific fight. A striker facing a strong wrestler
may shift to a more defensive, clinch-oriented game. These dynamic style state variables are the core edge
of the model and drive matchup interaction features.
02 / MATCHUP MATRIX
Style vs Style Interactions
Six fighting styles (Striker, Wrestler, BJJ, Muay Thai, Sniper, Mixed) create 36 unique matchup interaction
terms. Each interaction is learned independently by the model. A BJJ specialist against a pure striker
is treated as a fundamentally different contest than two strikers meeting. These interactions capture what
simple head-to-head stats cannot.
03 / DATA
10 Years of Fights
Scraped from UFCStats, every fight since 2015. Strike accuracy, grappling volume, round-by-round
aggression, physical attributes, rankings, and lag career statistics computed without data leakage.
04 / MODEL
Dynamic Logit + LASSO
Logistic regression with L1 regularization. LASSO automatically zeros out irrelevant features, keeping
only those with genuine predictive signal. Cross-validated lambda selection. Fully leak-free rolling stats.
05 / ODDS
Market Integration
DraftKings odds scraped Friday morning. Market implied probability enters the model directly.
We look for fighters where our model probability diverges from the market to identify mispriced bets.
06 / SIZING
Kelly Criterion Sizing
Continuous bet sizing via logistic-bounded Kelly. Scales with confidence and edge.
Minimum 1 unit. Maximum 5 units. Never recommends negative expected value bets.
07 / FINISH
Finish Type Model
Separate multinomial logit predicts KO/TKO vs Submission vs Decision probability for every fight.
Same LASSO approach. Visible as finish bars on each prediction card.
08 / PROPS
Prop Bet Screening
Strict threshold: edge greater than 15 percentage points AND model confidence above 0.70 AND
market probability below 0.65. Hidden by default and only shown when a genuine edge exists.
Disclaimer: OctaStats is a personal research project. Predictions are for entertainment and educational purposes only. Bet responsibly. Past model performance does not guarantee future results.