+11.92% median CLV · +1583% ¼-Kelly return · live every 5 min

Boxing prediction marketspriced by a model that beats the close.

Our model prices every upcoming boxing fight, tells you exactly which side to bet, at which sportsbook, and how much to stake. Every number is backed by an auditable record — no guesswork, no hand-waving.

+11.92%

Median CLV

892 gated picks

+1583%

¼-Kelly return

production gate · paper

5min

Refresh rate

odds + predictions

75

Sportsbooks tracked

across 5 regions

The model

A machine learning ensemble built for boxing markets.

The model that ships is the one that survived a rigorous, pre-registered audit. Every design decision — from feature selection to bet sizing — was validated on held-out data before going live.

25+ Features

Ratings · biometrics · form · market prices

rating_diff
win_pct
age
reach
market_p_a
rd_glicko

10× LightGBM

350 trees per seed · 3-class softmax

Seed 0
Seed 1
Seed 2
...
Seed 9

Vector Scaling

4 params · A/B symmetry by construction

scale_ab
shift_ab
scale_draw
shift_draw

3-Class Output

P(A) + P(B) + P(draw) = 1

P(A win) = 0.672
P(B win) = 0.291
P(draw) = 0.037

BET / NO BET

Edge ≥ 3% at best book → BET

Kelly stake
best book
email alert

How the model transforms data

10-seed ensemble

Each bar is one model's P(A wins). Hover to isolate a seed. Tight clustering = high confidence.

68.8
S0
66.4
S1
70.1
S2
68.0
S3
65.1
S4
69.2
S5
67.0
S6
68.5
S7
65.8
S8
69.5
S9
67.8%
MEAN

P(A win)

0.678

P(B win)

0.284

P(draw)

0.038

σ (uncertainty)

16.9 / 1000

SHAP feature waterfall

Starting at 35% base rate, each feature pushes the probability up (green) or down (red). Hover for details.

base rate
35%
rating_diff
recent_form_a
win_pct_career
market_p_a_open
rd_glicko_a
age_diff
sos_glicko
sched_rounds
reach_diff
other (16)
P(A wins)
67.2%
Calibration

Predicted probability vs actual win rate. Dashed line = perfect. Hover a bar. Toggle to see the correction.

100%75%50%25%0%
+11%
+9%
+6%
+9%
+9%
+11%
+9%
0–10%
10–20%
20–30%
30–40%
40–50%
50–60%
60–70%
70–80%
80–90%
predicted probability →avg error: 3.2%

Multi-seed ensemble

10 models trained independently, each with slightly different randomization. The average prediction is more stable and accurate than any single model — and the disagreement between seeds tells us how confident the pick is.

Calibrated probabilities

Each model is individually calibrated so the probabilities it outputs actually mean what they say. When the model says 65%, it wins ~65% of the time. This is essential for correct bet sizing.

3-class output with draws

The model predicts three outcomes — Fighter A win, Fighter B win, and Draw — learning draw probability from feature interactions. Evenly-rated fighters at high scheduled rounds produce higher draw probability than heavy favorites.

25+ features, zero data leakage

Fighter ratings, physical attributes, recent form, career stats, and opening market prices. Only data available before the fight starts is used — six structural safeguards prevent the model from ever seeing results or closing prices.

What the audit says

Audited performance. Every number verifiable.

Walk-forward backtest under our exact production parameters — the same gate that selects live picks: minimum 3% edge, reward-to-risk capped at 3-to-1, quarter-Kelly staking capped at 3% of bankroll, best price across all 75 tracked sportsbooks, no slippage. 892 qualifying picks. Every figure is computed live from the database with the same engine as the in-portal simulator.

Median CLV

+11.92%

production gate · n=892

Mean CLV

+25.35%

fat-tailed by winners

% beat the close

100%

share of gated picks with +CLV

¼-Kelly return

+1583%

892 picks · 3%/bet cap · edge ≥ 3%

Win rate

83%

gated picks, best book at close

Max drawdown

20%

¼-Kelly equity, peak-to-trough

These numbers are deliberately conservative

Backtests use the closing line — the tightest, hardest-to-beat price — as the bet price. In practice, you bet hours earlier when odds are softer. The reported numbers should be a floor, not a ceiling. Act fast when a BET appears.

Live betting page

Binary BET / NO BET. Stake $ at your bankroll.

No edge sliders, no opt-out toggles, no decisions for the user to make. The production policy decides which picks are bet-eligible. The calculator decides how much to bet. The card tells you the side, the best book, the price, and the stake. That's it.

  • Minimum 3% edge required — only picks with real value pass
  • Exact dollar stake computed for YOUR bankroll size
  • Best price across 75 sportsbooks, automatically selected
  • Picks auto-organized: upcoming, awaiting result, settled
  • Live freshness indicator — always know how current your data is
  • Email alerts the instant a new bet opportunity appears
  • Per-pick explainer — see exactly why the model made each call
See the live page
/portal/live
● Live data fresh · last tick 2m ago · next ~3m
BET2026-05-15 · predicted 4m ago

Karen Chukhadzhian vs Paddy Donovan

Pick: Karen Chukhadzhian @ 1.85 · best book pinnacle

Stake

$185

of bankroll

1.85%

payout if win

+$157

Edge

+8.20%

P(pick)

67.5%

σ

0.007

Kelly f*

7.40%

Bet Optimizer

Three tools beyond the BET / NO BET card.

Even when no fight meets the minimum edge threshold, sportsbooks still disagree on pricing — and those disagreements are opportunities. The Bet Optimizer catches arbitrage, positive-EV value, and hedging opportunities across all your books.

Arbitrage scanner

Finds pricing gaps between sportsbooks. When two books disagree enough, betting both sides at different books can return the same net on every outcome — a positive result only if both legs fill at the quoted prices.

Boxing price spreads last longer than in other sports — often hours, sometimes days. We flag the best execution order to reduce the risk of one side repricing before you place both legs. Execution and counterparty risk always apply.

+EV opportunities ranked by ROI

Every pick where the model sees a pricing edge, ranked by return per dollar. Some opportunities don't meet the strict BET threshold but still have value for more aggressive strategies.

A different lens than the BET cards. The +EV list surfaces everything the model likes ranked by capital efficiency, so you can size at your own discretion.

Hedge calculator

Already placed a bet? Enter your stake and the calculator shows how much to bet on the other side to even out your net across outcomes or cap a loss.

Useful when lines move after you've bet. A pick at +180 on Tuesday might shift to +110 by Friday — hedging can convert that move into a positive net if the hedge fills. Figures are modeled, not assured.

Tailored to YOUR sportsbook accounts.

We track prices from 75sportsbooks across five regions, but you don't have accounts at all of them. Select only the books you can actually bet at — arbs, +EV picks, and hedges are recomputed against just those. Your selection persists between sessions. Presets for sharp books, soft books, or regional groups.

The pipeline

From OddsAPI to your inbox in 5 minutes.

Odds are refreshed every 5 minutes. The model retrains overnight on the latest data. The entire cycle — from fresh odds to a BET card in your inbox — takes under 30 seconds.

  1. 1 · Ingest

    Every 5 minutes, fresh odds from 40+ sportsbooks are pulled and stored. Fight results are reconciled automatically from multiple data sources.

  2. 2 · Predict

    The model runs on every fight with current odds. Each prediction includes the ensemble's uncertainty estimate and a full SHAP breakdown of which factors drove the call.

  3. 3 · Filter

    An audit-derived policy decides BET or NO BET. Only picks with at least 3% edge pass. Bet size is computed via Kelly criterion and capped to protect your bankroll.

  4. 4 · Notify

    New BET-eligible picks trigger email alerts within seconds. You're never notified about the same fight twice, and the Bet Optimizer updates with fresh cross-book prices.

How we keep the model honest

The discipline that lets the numbers mean something.

Test data sealed before modeling began

The holdout set was locked before we wrote a single line of model code. Results were only checked once, after the full audit.

Six anti-leakage safeguards

Structural protections ensure the model never accidentally sees fight results, closing prices, or future data during training.

Every number is auditable

All reported metrics live in the database, not in code. The portal reads them directly — nothing is calculated on the fly or derived client-side.

Daily re-audit catches drift

The full pipeline re-runs every night. If the model's performance degrades, you see it the next morning — no lag, no surprises.

Deliberately conservative reporting

Backtests use the closing line — the tightest price available. Real execution at earlier, softer prices should outperform the reported numbers.

Full development history in the portal

Every experiment, negative result, and design decision is documented in the System Development Process tab. Nothing is hidden.

How it works

Three steps. One engine.

01

Open your portal

Create an account and you're in. You'll land on the live predictions page immediately.

02

Set bankroll, pick your books

Enter your bankroll once. Every BET card shows the exact dollar stake and payout at the best-priced sportsbook. Select only the books you have accounts at.

03

Get emailed when bets land

Turn on email alerts in settings. New bets hit your inbox within seconds of detection — so you can act before the line moves.

The model is the boring part. The discipline is what matters. Test data sealed before we wrote a single line of code. Backtests use the hardest-to-beat prices. Every number independently verifiable. That's what lets the numbers mean something.

— operating principle

FAQ

Common questions.

Live now · refreshes every 5 min

See the bets.See the math behind every one.

Live bet cards with exact stakes, cross-book arbitrage and +EV tools tailored to your sportsbooks, a full audit dashboard, and instant email alerts.