Full Transparency

Model Configs & Validation

Every parameter, every threshold, every p-value. Nothing hidden. These are the exact configs running in production.

Research completed 2026-03-21 · ~500,000 configs tested · 3 sports validated

Combined Portfolio

+17.5%

Backtest ROI

weighted avg

738

Total Bets

~66/year

Bets / Year

3 sports combined

p < 0.01

All Significant

permutation test

Boxing

0 negative years

Favorites — XGBoost no-form + Elo

Backtest Results

+18.1%

Flat ROI

+10%

Holdout ROI

out-of-sample

60.2%

Win Rate

0.0000

p-value

permutation test

251

Total Bets

12 years

Bets / Year

0/12

Neg Years

12 years

Period

Model Configuration

model:XGBoost 5-seed ensemble

features:28 (no-form set)

regularization:L2 = 10

calibration:Isotonic regression, 20% held-out

augmentation:Fighter A/B swap for balanced labels

retrain:Every 120 days

elo K:K = 32

Selection Filters

Min Edge≥ 12.5%

Elo Prob≥ 40%

Odds Range1.20 – 4.50

KellyQuarter-Kelly, max 10%

Feature Groups (28 (no-form set) total)

Physical3

height, reach, age advantage

Elo3

ratings + differential

Fight History6

counts, win rates, KO rates

Boxing-Specific8

decision rate, avg rounds, KO loss rate, activity

Quality Proxy3

career avg odds, title fight

Market3

implied probabilities + diff

Dropped: streak_a, streak_b, recent_winrate_diff (form features removed — reduced noise without losing signal)

Permutation test: 5,000 permutations of model signal. Zero permutations beat the actual model.

Soccer Draw

Highest ROI per bet

Draw predictions — EPL + Ligue 1 only

Backtest Results

+24.1%

Flat ROI

+25.8%

Holdout ROI

out-of-sample

35.4%

Win Rate

0.008

p-value

permutation test

195

Total Bets

19 years

Bets / Year

2/19

Neg Years

19 years

Period

Model Configuration

model:XGBoost 5-seed ensemble, 3-class (H/D/A)

features:20

regularization:L2 = 10

calibration:Isotonic regression, 20% held-out

augmentation:Home/Away swap (label: H↔2, D↔D, A↔0)

retrain:Every 120 days

elo K:K = 48

Selection Filters

Min Draw Edge≥ 10%

Min Draw Prob≥ 28%

Max Draw Odds≤ 3.50

Max Elo Home≤ 58%

LeaguesEPL + Ligue 1 only

KellyQuarter-Kelly, max 10%

Feature Groups (20 total)

Elo3

home, away, differential

Attack/Defense6

goals, conceded, shots on target

Home/Away Split2

win rates by venue

Match Stats4

corners, fouls averages

Season Context2

points per game

Market3

implied probabilities (H/D/A)

Permutation test: 5,000 permutations of model signal. Only 0.8% of permutations beat the actual model (p = 0.008).

Note: Draw signal is league-specific. Tested on 11 additional leagues: −5.93% ROI. Only EPL and Ligue 1 show exploitable draw mispricing.

UFC

Highest volume

Favorites — XGBoost + Elo consensus

Backtest Results

+12.5%

Flat ROI

+36.4%

Holdout ROI

out-of-sample

66.8%

Win Rate

0.003

p-value

permutation test

292

Total Bets

9 years

Bets / Year

2/9

Neg Years

9 years

Period

Model Configuration

model:XGBoost 5-seed ensemble

features:33

regularization:L2 = 10

calibration:Isotonic regression, 20% held-out

augmentation:Fighter A/B swap for balanced labels

retrain:Every 90 days

elo K:K = 32

Selection Filters

Min Edge≥ 5%

Max Odds< 2.0 (favorites only)

Elo AgreementElo prob ≥ 40%

KellyQuarter-Kelly, max 10%

Feature Groups (33 total)

Physical3

height, reach, age advantage

Elo3

ratings + differential

Fight History13

counts, win/KO/sub rates, streak, form

Fight Stats14

sig strikes, TD, control, knockdowns per fight

Permutation test: 5,000 permutations of model signal. Only 0.3% beat the actual model (p = 0.003). Replicated across 5 L2 values.

Note: 33 features outperform 54 features (bias-variance tradeoff). Adding implied_prob destroys the signal (−1.2% ROI) — the edge comes from non-market information.

What Didn't Work

Transparency means showing failures too. These sports/strategies were tested exhaustively and rejected.

×NBAMarket too efficient

22K+ config sweeps, Elo, XGB, dynamic home advantage, rest days — all negative or near-zero holdout.

×NHLInsufficient signal

Permutation p = 0.44. Best year driven by single outlier (+197%).

×TennisMarket too efficient

−8% to −19% ROI across all configs. Well-calibrated but can’t overcome 5–7% bookmaker margin.

×Extra Soccer LeaguesDoesn’t generalize

Draw strategy on 11 new leagues: −5.93% ROI, p = 0.75.

×Multi-Book AlphaNot significant

Line movement p = 0.10–0.20. Sharp/Soft divergence not reproduced on retest.

Validation Framework

Every strategy must pass ALL four gates before going live:

1. Holdout Positive

Train on data before cutoff date, test on remainder. Must show positive ROI on unseen data.

2. Permutation p < 0.05

Shuffle model signal across all matches, re-apply filter. Tests if model selection beats random.

3. Pre-Registration

Hypothesis stated before seeing results. Max 5 configs per target to prevent data mining.

4. Replication

Positive ROI across multiple regularization values (not dependent on a single hyperparameter).

Questions about our methodology? Read the full methodology guide