Measured · not projected

What Antler finds, and how we know

Two questions. Every figure below is measured, not projected.

+1.18%

+EV return · 461 bets

+1.32%

Arbitrage · 274 positions

+1.35

Median CLV · points

72%

Bets beating the close

I am a

Part one

What it finds

A bookmaker's price is a prediction dressed up as a number, and different bookmakers make different predictions about the same game.

Most of what we find works like this. A handful of bookmakers are genuinely excellent at pricing, because professionals bet into them and correct their mistakes. Most are not. When a sloppy one offers a better price than the excellent one says is fair, that gap is worth money — on average, across hundreds of bets. Never on any single one.

Expect to lose plenty. Our own record is 210 wins against 244 losses, and still profitable, because the winners pay more than the losers cost.

The second kind is not a prediction at all. If one shop is unusually generous about the home team and another about the away team, you can back both sides and finish ahead whichever one wins. That is arithmetic, not judgement. It is also small, and it disappears fast — about a quarter are gone within a minute.

Part two

How we know it works

If a tool makes money over a few hundred bets, that could be skill or it could be luck. There is no way to tell them apart from the total alone.

So we test something else. In the minutes before a game starts, the price stops moving. Everyone who was going to bet has bet, including the professionals, and that final price is the closest thing there is to the truth about the game.

The test is simple: did we get a better price than that final one? If we did, we found something real — and that stays true whether the bet won or lost, because a good price on a losing bet is still a good price.

Right now 7 out of 10 of our bets beat the final price. That is the number that says this works, and it would still say so in a week where we lost money.

Part one

What it finds

Two products, one pipeline. They behave completely differently and are never reported as one number.

Positive EV. We take Pinnacle's price, strip out its margin to get a fair probability, then look for a soft book paying more than that. A typical edge is 2–4%; we are running +1.18% over 461 bets. You lose most of them — 210 wins against 244 losses — and still profit, because the winners pay better than even.

The benchmark is never an average of DraftKings and FanDuel. Soft books copy each other, so their consensus inherits the same bias.

Arbitrage. Best price on each side across every book we watch. When the implied probabilities sum under 100%, staking both sides in the right proportion locks in profit regardless of result. +1.32% ROI over 274 settled positions — guaranteed on paper; in practice a leg gets rejected or the price moves while you are placing the second.

Part two

How we know it works

Closing line value: how much better your price was than the fair price at close, measured in probability points.

Why not just show profit? Each +EV bet swings roughly ±$100 against an expected edge of a few dollars. Realised P&L needs thousands of bets to separate skill from noise. CLV converges far faster, which is why it is the standard measure of bet quality.

Current reading: median +1.35 points across 461 settled bets, with 332 of them positive.

Only bets you could actually have placed count. We re-check every alert about a minute after sending it — a price already gone says nothing about the quality of the advice, only about speed.

Part one

What it finds

Two detectors. Only one of them requires a probability estimate.

EV = p_fair(d − 1) − (1 − p_fair)

p_fair comes from the sharp reference, devigged. Four methods are implemented — multiplicative, additive, power, Shin — and the method used is recorded on every row, because on a heavy favourite the inter-method spread reaches 2.57 probability points. That is the same order of magnitude as the edge being claimed, so an edge only one method finds is not displayed as an edge. Cross-method agreement ships as a confidence indicator on every card.

Sizing is fractional Kelly. The shadow book is flat-staked on purpose, so results never depend on an arbitrary starting bankroll.

arb: Σ_i 1 / max_b(d_i,b) < 1

Deterministic and model-free — no p̂, no calibration risk. Stakes fall out of the reciprocals. Exchange prices are taken net of commission and quoted at the ask, never the mid.

Part two

How we know it works

CLV = p_fair,close − implied(d_taken) [points]

Points, not a ratio. EV is linear in probability, so a ratio distorts across the range — 5%→6% and 50%→60% would read identically.

Entry and close are devigged with the same method, or the statistic measures method disagreement rather than movement in the line.

Population: persisted_60s = 1 only — confirmed still available at the advised book 60 seconds after being flagged. An opportunity gone before anyone could act says nothing about the quality of the advice.

Never condition on a quantity sharing a term with the metric: splitting by whether the line moved selects on p_close, which sits inside CLV.

n=461 · median +1.35 · mean +1.16 · 72.0% positive. The pure line-movement measure is kept separately as clv_move(), useful precisely because it is independent of what we claimed.

What would change our mind

A tool that only reports its good weeks is worse than no tool. These are the live limits, stated at the same volume as the results above.

The sample is still young

461 settled +EV bets is enough for CLV to speak and not enough for profit to. Treat the return as descriptive and the CLV figure as the evidence.

The edge is not in the big markets

Major-market sides and totals are efficient everywhere. Mispricing lives in props and secondary markets, where soft books price algorithmically and do not sharpen until money arrives.

Speed is not our advantage

+EV prices barely move — only 1% are gone within a minute of being flagged. Arbitrages are the opposite at 24%, and we do not compete on latency.

Arbitrage always flatters itself

269 wins, 0 losses is what arbitrage looks like on paper by construction. Real execution rejects a leg, moves the price mid-placement, or voids the bet — leaving one side exposed.

Measures the disagreement.