Proof & records
Methodology
How predictions are made, how picks are graded, and where to verify the record.
Bet responsibly.
Predictions are not financial advice. Past accuracy does not guarantee future results. If gambling is becoming a problem, call or text 1-800-MY-RESET.
Refreshing
Graded Public Hit Rate
0
Graded Public Picks
Withdrawn
Walk-Forward Benchmark
n/a
Overall Brier Score
15,000
Monte Carlo Sims / Fight
Accuracy claims come from graded public picks scored against real results. The walk-forward backtest is research, not a benchmark, and is labelled that way
on the Accuracy page.
Receipts & Model Q&A
Verify the record, then ask how to read it
The prediction archive shows scored post-event receipts. Ask the Model is a read-only explainer that summarizes stored FightEdge evidence without generating a new pick.
How a Prediction Is Made — Step by Step
1
Data Ingestion
Multi-source UFC fight data — UFC.com (primary), Wikipedia, Sherdog, plus per-round stats from 6,871 historical fights. Live odds from multiple sportsbooks. Expert picks from MMA Intel and DRatings. Reddit community sentiment.
2
Fighter Profile Build
Each fighter's profile is computed from their UFC + career fight history: striking accuracy, defense, takedown defense, control time, durability, finish rates by method, archetype classification (striker / wrestler / grappler / well-rounded), and Elo rating with style-matchup adjustments.
3
Rule + Elo Predictor
A rule-based predictor combines the Elo head-to-head delta with stylistic matchup factors (e.g. wrestler vs striker, range advantage, durability gap). The graded public record is refreshing; accuracy claims are withheld until the current verified snapshot is ready. ML pipeline retired in v3.2.7 after the same walk-forward framework proved it was Brier-worse than the rule baseline.
4
Monte Carlo Simulation
15,000 fight simulations per matchup. Round-by-round striking, grappling, fatigue, damage accumulation, dynamic chin degradation, stance matchup modifiers, division-specific finish rates. Generates method probabilities, fight-duration props, per-zone strike projections, takedowns, control time, round-by-round finish probability.
5
Isotonic Calibration
Probabilities are calibrated using isotonic regression trained on the rule+Elo predictor's own walk-forward output. Brier improvement: 0.0244. Calibration table is fitted on 2,506 training points and validated on 627.
6
Output
Win probability, method breakdown (KO/TKO, SUB, DEC), fight duration props with fair odds, per-round finish probability, per-zone strike projections, takedowns, control time. Plus signal attribution showing which inputs drove the prediction. Plus fully-AI-written 3-section fight breakdown via Claude Opus.
Calibration Detail
Confidence labels (FLIP / LEAN / SOLID / STRONG) are calibrated using isotonic regression so that a "70% pick"
actually wins ~70% of the time historically. The calibration layer is fitted on the rule+Elo predictor's own
walk-forward output and capped at 0.85 to respect the variance UFC fights actually exhibit.
Method
Isotonic regression
Training Set
2,506 fights
Brier Improvement
−0.0244
Calibrated Holdout Brier
0.1801
Live calibration curve and per-bucket hit rates are tracked on the
Calibration page.
Transparency Comparison
Most UFC prediction sites advertise high accuracy with no public verification path. Here's what FightEdge publishes
versus what's typical in the space.
| What We Publish | FightEdge | Typical Prediction Site |
|---|---|---|
| Walk-forward accuracy as a public benchmark | Withdrawn | Not typical |
| Per-window walk-forward breakdown (research, Accuracy page) | Published | Not typical |
| Brier score (overall + holdout) | Published | Not typical |
| Data sources: UFC.com, Wikipedia, Sherdog, and multi-book odds | Published | Not typical |
| Validation cadence and source-count freshness checks | Published | Not typical |
| Walk-forward research backtest, labelled, on the Accuracy page | Published | Not typical |
| Calibration methodology + training-set size | Published | Not typical |
| Retraining policy and scheduler visibility | Published | Not typical |
| Live per-event accuracy tracking | Published | Not typical |
| Public calibration curve | Published | Not typical |
| Closing Line Value (CLV) tracking per pick | Published | Not typical |
| Sample-size buckets / data-depth labels | Published | Not typical |
| Method probabilities (KO/SUB/DEC) per fighter | Published | Available |
| Win probability percentages | Published | Available |
| Confidence band labels (FLIP/LEAN/SOLID/STRONG) | Published | Not typical |
Verify Every Number Yourself
Live confidence-bucket calibration. See whether 70% picks really win 70% of the time.
Every prediction we've ever logged with its actual outcome. Per-event breakdown.
Live community-vs-model accuracy, with the model pinned as its own row.
Full platform story, what we built, who it's for, and how the data flows end-to-end.
Practical Trust Guide
Reading Model Confidence
Confidence zones are probability bands, not guarantees. The Calibration page shows how each public zone has actually performed after fights score.
CLV Tracking
Closing-line value compares the price you logged to the closing market for that picked side. It measures whether your process beat the market, even before a single fight outcome is known.
Kelly & Sizing
Any sizing guidance should be treated as educational. Real-money bankroll decisions need conservative caps, skipped thin edges, and variance tolerance.
Paper Trading
Paper bets are a separate no-money practice ledger. They auto-score separately and never affect real ROI, bankroll history, CLV journal, or badges.
Adaptive Personalization
Adaptive readiness is display-only. It summarizes your graded real-bet sample and signal-arm history; it does not rewrite stored probabilities or change public model picks.
Truth contract
Every customer-visible number on FightEdge has exactly one producer, and pages display that producer's value instead of recomputing it. The rules we hold ourselves to:
- Picks, win probabilities, and method leans come from the sealed analysis record for that fight, never from a page's own math.
- Results and settled counts come from the graded ledger, and a page never mixes counts from different sources without saying so.
- Empty states say why they are empty; nothing renders a stale figure as current.
- Opening a page never triggers scraping, analysis, or model work; you see certified data, not a live computation.
- Presentation flags change layout only; they can never hide, alter, or gate a number.
- Accuracy claims come from the graded public record; the walk-forward backtest is research, labelled on the Accuracy page, and is not used as a benchmark.
A note on what these numbers mean:
A publishable high hit rate on winner picks is exceptional in MMA — UFC fights are inherently variable, and elite handicappers
typically operate in the 60-65% range. But "high accuracy" alone doesn't guarantee profitable betting; profitability
requires accuracy AND finding edges where the market is mispriced. That's why we also track Closing Line Value
(CLV) per pick — see the Accuracy page. We publish all of this so you can
decide for yourself, not because we're guaranteeing outcomes. Predictions are not financial advice. Bet responsibly —
call or text 1-800-MY-RESET if you need help.