Engine 4.0 System
A stacked machine-learning ensemble that turns 6,000+ historical fights into a calibrated win probability — and an honest bet signal — for every matchup.
What Engine 4.0 is
Not one model, but several. Engine 4.0 runs two gradient-boosting models ( XGBoost and LightGBM) and blends them with a stacking meta-learner, then calibrates the result so the percentage you see is the percentage it actually hits.
From raw data to a pick
Every prediction travels the same six steps.
Fight data
6,000+ historical bouts with per-fighter stats, odds, and outcomes feed every prediction.
Feature engineering
Raw stats become matchup features: ELO gaps, form, layoff, striking/grappling differentials, physical edges.
Base models
Gradient-boosted trees (XGBoost, LightGBM) each score the matchup independently.
Stacked ensemble
A meta-learner combines the base models' outputs into one blended probability.
Calibration
Platt scaling maps raw scores to honest probabilities — a 65% means it hits ~65% of the time.
Bet signal
The calibrated probability and the edge vs Vegas decide the tier: STRONG BET, BET, LEAN, or SKIP.
The ensemble
Different models make different mistakes. Blending them cancels out individual blind spots and beats any single model on its own.
XGBoost
Gradient-boosted decision trees that capture non-linear interactions between matchup features.
LightGBM
A faster histogram-based booster that complements XGBoost with different split behaviour.
Stacking meta-learner
Learns how much to trust each base model per matchup, then outputs the final blended probability.
Calibrated, not just confident. Raw model scores are passed through Platt scaling so the probabilities are honest. Training also uses mirror augmentation — every fight is learned from both corners — so the model never inherits a red-vs-blue bias.
What it looks at
Features fall into four families. For a deeper, effect-size ranking of what actually moves fights, see what predicts fights.
ELO ratings & momentum, age, win/loss records, recent form, layoff
Strikes landed/min, accuracy, absorbed, striker rating
Takedown average & accuracy, takedown defense, sub attempts, control
Title bouts, rankings, reach/height/weight differentials
How it got here
Four generations, each adding signal and refining the architecture.
Engine 1.0 — Foundation
Basic fighter stats, records, and physical edges — the first working prediction engine.
Engine 2.0 — Feature Expansion
Added ELO ratings, recent-form metrics, layoff analysis, and striker-vs-grappler classification.
Engine 3.0 — Ensemble Architecture
Multi-model stacking, probability calibration, mirror augmentation, and finish-method prediction.
Engine 4.0 — Current — Refined Ensemble
ActiveXGBoost + LightGBM + stacking with calibrated probabilities, tuned on the full historical dataset.
See the engine in action
Every upcoming fight, scored and tiered.