The Method

How a pick
gets made

Six steps from raw numbers to a graded call.

01Read the market
02Rate the fighters
03Two AIs make the call
04Anchor to the market
05Grade every pick
06Price a trade

Read the market

The betting line is the starting point. It is the sharpest public number there is.

p = 100 / (odds + 100)

underdog, odds above 0

p = |odds| / (|odds| + 100)

favorite, odds below 0

p₁′ = p₁ / (p₁ + p₂)

remove the bookmaker's margin so both sides sum to 100%

Worked example

Fighter A -20066.7%

Fighter B +17037.0%

Adds up to103.7%

Fair line64.3% / 35.7%

With no sportsbook line we use Kalshi's exchange price the same way.

Rate the fighters

Every UFC fight since 1994 feeds one rating. Beating good fighters counts more than beating weak ones.

E_A = 1 / (1 + 10^{(R_B − R_A) / 400})

expected chance A beats B

R_A′ = R_A + K · (S − E_A)

S is 1 for a win, 0 for a loss

Start at 1500K = 2440 under 5 UFC fights×1.2 for a finish

Two AIs make the call

Claude and GPT get the same file on every bout and pick independently.

UFCStats career numbersLast 5 fights and layoffElo ratingsMarket lineLine movementFight-week news with sources

Claude

GPT

Out

WinnerWin chance, 50 to 90%KO, sub or decision splitKey factors

61.5%when both agree, 569 fights

72.5%double locks, both 75%+, 51 fights

Anchor to the market

The models are good, the market is better. So the stored confidence leans on the market. The pick itself never changes.

logit(p) = 0.3 · logit(p_model) + 0.7 · logit(p_market)

logit(p) = ln(p / (1 − p))

shown = max(50%, p)

a pick the market dislikes shows as a coin flip

Brier score on 121 priced bouts. Lower is better.

Try it

Model says 65% Market says 55%

Shown on the site58%

Grade every pick

Right or wrong is not enough. We score how honest each number was.

Brier = (1/n) Σ (p − o)²

p is the stated chance, o is 1 if the pick won; always saying 50% scores 0.250

60.5%Claude right, 703 fights

57.4%GPT right, 751 fights

Market score is on the 121 bouts with a stored line.

ClaudeGPTPerfect

Below the dashed line means overconfident. Dot size is sample size.

Price a trade

A Kalshi contract pays $1 if the pick wins. It is only worth buying above break-even.

fee = 0.07 · P · (1 − P)

per contract, P = price in dollars

break-even = P + fee

EV = p − P − fee

per $1 contract, p = true win chance

f = (p · b − (1 − p)) / b, b = (1 − P) / P

Kelly stake; edge sizing bets a fraction of this

What the record says

Model picks the favorite

Won68%

Price paid67%

102 picks

Model picks the underdog

Won37%

Price paid39%

19 picks

Gold tick is the price paid. Past results are small samples, not a promise.

Research lab: audited walk-forward results

No proven trading edge. Use paper tracking while collecting timestamped prices, fills and actual fees.

Model	Bouts	Brier	Market
Sportsbook benchmark	53	0.246	0.246
Market + strongly shrunk model signals	53	0.246	0.246
Market + model signals	53	0.248	0.246
Market + lightly shrunk model signals	53	0.249	0.246
Model chosen using earlier events only	53	0.246	0.246

Stats model, untouched test set	Brier	Accuracy
logistic_C0.1	0.239	58.0%
logistic_C1	0.239	58.0%
forest	0.239	59.0%
boosting	0.240	58.0%
elo	0.246	55.0%

Trained on earlier cards only, tested on the next one. Updated 2026-11-04T02:57:13.363Z.

Prediction accuracy, sportsbook simulations and real Kalshi trades are tracked separately. See the record