Need A Birdie AI Method report
How the number is made, and how the record is kept honest.
We publish two things: a probability the market has not fully priced, and a record you can audit without trusting us. This page explains the machinery in plain English. The proof lives on the dark pages.
Explore the method13 chapters
What Apollo does
Apollo is the engine behind every pick on this site. It does five things, in a fixed order, every day of a tournament week: it prices golf with its own simulation, it reads what the market believes, it blends the two and charges itself for doubt, it publishes only what survives, and it seals what it published so the record cannot be argued with later.
Nothing about that is magic. It is disciplined analytical machinery, and every step leaves a row a stranger can check.
From a price to a probability
A sportsbook price is a probability wearing a margin. Add up the implied chances of both sides of any market and you get more than 100 percent; the excess is the book’s cut, the vig. Before we compare our number to the market’s, that margin has to come out, and how you take it out matters.
| Side | Price | Implied | No-vig |
|---|---|---|---|
| Player A | +140 | 41.7% | 40.1% |
| Player B | −165 | 62.3% | 59.9% |
| Total | 103.9% | 100.0% |
The 3.9 points above 100 are the book’s margin. Removing it proportionally gives the market’s honest opinion of each side.
The technical disclosure
Splitting the margin evenly is only safe on balanced two-way markets. On longshot-heavy boards it manufactures fake value on the long prices, which is exactly where a golf model most wants to fool you. We remove the margin with methods suited to the market’s shape: one approach for longshot-heavy boards, a simpler one for two-way markets. There is one implementation, and everything calls it.
Our own number
The model prices golf by simulating the tournament ahead of it, many thousands of times, from data that existed at the moment of the bet. The simulator does not treat golfers as independent coin flips: wind, waves and course conditions hit whole groups of players together, so the draws move together too. That correlation is most of the reason a field’s variance looks the way it does, and pricing it honestly is most of the work.
A player’s own record at a venue and his recent form are inside the number. Every input is provenance-tagged, so a number we publish can be audited back to its sources.
The technical disclosure
We tested course-type fit, the idea that certain styles of player suit certain styles of course, and it made our forecasts worse, so it carries zero weight. What survives is the player’s own venue history, measured over his actual rounds there. No feature may use information that did not exist at the moment of the bet; every new feature ships with a point-in-time test proving it.
The blend
The market gets a vote.
Our raw number is blended with the devigged market consensus, weighted heavily toward the market, and a play only makes the card when the disagreement survives that blend. The market is the sharpest opponent we have; treating it as noise would be the fastest way to look brilliant and lose money. Most disagreements do not survive. The ones that fail but came close are published anyway on the watch list, graded and never staked, so the near misses have a scoreboard too.
Why a big edge can still be a No Play
A large raw disagreement is an argument, not a bet. After the blend pulls our number toward the market and the doubt charge is paid, what is left has to carry a full unit of stake. When it cannot, the pick is published as a No Play: sealed, graded and counted for calibration, never staked. This is the honest middle between hiding the pick and pretending it deserved money.
You will sometimes see a No Play wearing a double-digit raw edge. That is the system working, not failing: the bigger the disagreement with an efficient market, the more of it is usually our error, and the doubt charge prices that.
The stake
Plays are sized by the edge left after doubt, in units: tier A is two units, tier B one and a half, tier C one. Nothing is ever staked below one unit; if a play cannot carry a full unit it is published as a No Play. Every play is quoted at a major US sportsbook’s own posted price, and the sealed row records the price we actually took.
The seal
Every play is written to the ledger before it can be judged: the price taken, the book, the stake, our probability and the market’s at that moment, and the reasoning, all recorded before the player’s own pairing tees off. The database physically rejects edits and deletes to sealed rows. A better board before the gun is a new revision, a fixed mistake is a new row that says who fixed it and when, and the old rows stand forever.
That is the whole trick. There is no way to look good retroactively, so the record means something.
The scoreboard
The first number fans see on every results report is units won or lost. Alongside it, we show drift: how far the closing consensus moved toward the sides we sealed. A number ahead of the close tends to make money over time; a number behind it tends to give it back, whatever last week’s results said. Units lead the public scoreboard because that is how fans track betting results. Drift remains visible because it is how we judge the model.
The track record shows every real-money play and its official result. The Full ledger holds every sealed row behind it, including No Plays, corrections and superseded cards.
The humans
People in this group have played and caddied tournament golf for decades, and what they see is welcome, within a fence: a human read enters the model as a small, capped adjustment with an author, a timestamp and a confidence, and every read is scored against the outcome afterward. Nobody, including us, can reach in and move a number by hand.
The limits of the model
An honest method chapter ends with what can beat us, so here it is, plainly.
Variance beats us often. Golf is the highest-variance major sport there is; one round settles many of our markets, and edges our size lose all the time without the number being wrong. Samples are small. A few weeks of plays proves almost nothing in either direction, which is why units alone never tell us whether the model is sound. The market is good. Closing prices embed information we do not have, including money we never see. Weather is a forecast.The draw’s advantage is real but our daily read of it captures only part. Late news is late. A withdrawal or an injury after the seal belongs to variance, not to the model. And some ideas we hoped would help, like course-type fit, tested out as noise and were removed.
A pass counts.
What would earn stronger claims
Today we claim a disciplined process and a sealed record, nothing more. Stronger claims have a price, and the price is evidence: hundreds of staked closes with drift persistently on our side, calibration curves that hold at every probability we quote, and more than one season of out-of-sample weeks. Until the ledger contains that evidence, the modest claim is the true one, and this page will say so.
What this is not
Nothing on this site is betting advice. A pass counts. When a price is gone, it is gone, and no page here will ever encourage you to chase a worse number. Where you can bet, and whether you should, is yours to decide under the laws where you live. If betting stops being entertainment, help exists: in the United States, call or text 1-800-GAMBLER.
The words, precisely
- Probability
- Our model’s chance that a side wins its market, as a percentage.
- No-vig probability
- The market’s chance for the same side once the book’s margin is removed. The honest version of a price.
- Edge
- Our probability minus the no-vig market probability, in percentage points. A disagreement, not a promise.
- Expected value
- What a one-unit stake would earn on average at the taken price if our probability is right, after the blend’s doubt.
- Stake, unit
- The size of a play, in units of a fixed bankroll fraction. Tier A is 2u, B is 1.5u, C is 1u; nothing is staked below a unit.
- CLV
- Closing line value: how the price we took compares with the closing price. The classic test of whether a bettor is ahead of the market.
- Drift
- Our probability-points version of the same idea: how far the closing consensus moved toward or away from our side after the seal.
- Uncertainty
- The honest width around any number here. Small samples and high variance mean every figure on this site carries it.
- Calibration
- Whether things we call 40 percent happen about 40 percent of the time, measured across many plays. The model’s real exam, and it takes volume to grade.