How it works
7 LAYERS
OF INTELLIGENCE
01
01 · ML
MACHINE LEARNING
LightGBM
The model is trained on dozens of features: fighter form, opposition strength, physical parameters, style and matchup, PageRank graph metrics.
Momentum & recent formStrength of SchedulePhysical parametersStyle matchupPageRank metrics
Finds hidden patterns impossible to define by hand.
02
02 · ELO
ELO RATING
Dynamic strength
A classic rating system that accounts for opponent strength. Updated after every fight, it provides a stable strength baseline.
Opponent strength awareReal-time updatesStable rating baselineTime decay
The system's anchor, protecting against noisy data.
03
03 · GRAPH
OPPONENT STRENGTH GRAPH
PageRank-like algorithm
Every fighter is a node in a graph. Edges account for wins/losses, the type of win (domination or a battle) and how long ago the fight happened.
Graph of all fightersWin weight (domination / battle)Time decay on edgesIterative PageRank
A PageRank analogue — beating the strong makes you stronger.
04
04 · STYLE
STYLE MATCHUP MATRIX
Style compatibility matrix
The system analyzes style interactions: Top Roll vs Hook, Kings Move vs Press, Side Pressure vs Flop Press.
Top Roll vs HookKings Move vs PressSide Pressure vs FlopHistorical win rates
Predicts not just who is stronger, but how exactly the fight will unfold.
05
05 · LOGIT
LOGIT FUSION
ML + Elo in logit space
ML and Elo are merged in logit space — the model's confidence is preserved and probability dilution is avoided.
Logit spaceConfidence preservedWeighted fusionAnti-dilution
More honest and stable forecasts than simple averaging.
06
06 · UPSET
UPSET RISK ENGINE
Unexpected result detector
A separate layer estimates upset probability: lack of experience, unstable form, weak opposition, model disagreement.
Fighter experienceForm stabilityOpposition qualityML vs Elo divergence
Shows how risky the forecast is.
07
07 · CALIB
AUTO CALIBRATION
Confidence self-tuning
The system automatically adjusts its confidence: reduces favorite overestimation and adapts to real results.
Overconfidence reductionAdapting to resultsCalibration curvesPlatt scaling
Brings probabilities closer to reality.
Advantages
WHAT MAKES
M.I.R.A.
UNIQUE
Classic rating systems look only at the result. M.I.R.A. sees the context — who, against whom, with what style and in what form.
75-80%
Prediction accuracy
01
Style-aware
Analyzes fighting style, not just the winner's numbers
✓
02
Opposition strength
An opponent graph via a PageRank-like algorithm
✓
03
Hybrid model
ML and ratings merged in logit space
✓
04
Upset detector
Estimates the risk of an unexpected result
✓
05
AI explanation
The system explains its forecast in a fighter's language
✓
06
Auto-calibration
Probabilities adapt to real outcomes
✓
Pipeline
FROM DATA
TO FORECAST
Result
FINAL FORECAST + UPSET SCORE
→
Roadmap
THE FUTURE
OF THE SYSTEM
M.I.R.A. keeps evolving. Upcoming versions will add new dimensions of analysis.