Why We Have Not Shipped a Golf Board

The usual way to write about a missing feature is to promise it. We would rather explain it, because the reason golf is absent from our boards says more about how we build than any of the sports we do cover would.

The short version is that the engine is not the problem. A golf engine exists in our codebase, it simulates play hole by hole, it treats the cut as a structural event rather than a footnote, and it has been through an adversarial audit like every other engine we run. What it does not have is the data that would make its output mean anything. That is a procurement problem, not a calendar problem, and it is worth being precise about the difference.

Season gated and data gated are different states

We use two words for engines that are built but not live, and they describe genuinely different situations. A season gated engine is waiting for a clock. Football, the NBA and college football all waited in that state until their seasons produced real games and platforms posted real lines, and college basketball waits there now. Nothing is missing except time.

A data gated engine is waiting for an input. Golf is played somewhere nearly every week of the year, so no calendar is holding it back. What is holding it back is that we do not have a licensed, reliable, historically deep feed of the shot level skill measurements and tournament results that a credible golf projection is built on. We could open a board tomorrow. We would just be publishing numbers with far less behind them than the numbers on our other boards, and a reader has no way to see that difference from the outside. That is the whole reason we do not.

Why golf is unusually data hungry

Golf props look simple. A player shoots a score, makes or misses a cut, records some number of birdies. The simplicity is misleading, because producing an honest distribution over those outcomes requires more inputs than most team sports do, for four separate reasons.

The first is that golf performance decomposes into distinct skills that behave differently and stabilize at different speeds. Strokes gained is the framework the sport uses to separate them: how many strokes a player gains or loses relative to a field baseline off the tee, on approach, around the green, and putting. Those four are not interchangeable. Approach play is comparatively persistent from event to event; putting is famously noisy and regresses hard. A model that only knows a player's finishing positions cannot tell a good ball striker who putted badly from a mediocre one who putted well, and those two players will produce very different scores next week on a course that rewards approach play. Getting that separation right requires shot level measurement, not leaderboards.

The second is field size. A team sport slate gives you a few dozen games and clear pairings. A full field golf event puts well over a hundred players in the same competition, all of whom are being compared to the same baseline and to one another. Every player in the field needs a credible skill estimate, including the ones who qualified this week and have thin recent records, because the field baseline itself depends on who showed up. Sparse data on the back half of a field is not a rounding error. It moves the reference everyone else is measured against.

The third is the cut. In a standard event, roughly half the field goes home after two rounds under a cut rule, and everything about a player's tournament changes at that boundary. Whether he plays thirty six holes or seventy two is not independent of how well he played, so any prop scoped to the whole event carries a structural break in the middle of it. Our engine handles this by simulating each player's own first two rounds against a drawn cut line rather than applying an average, which is the correct approach and also the one most demanding of accurate skill inputs, because the estimate decides which side of the line the player lands on.

The fourth is that two players in the same event are not playing the same tournament. Tee waves split the field between morning and afternoon, and conditions move. Wind builds, greens firm up, a front arrives at noon. Two competitors with identical skill can face materially different courses on the same day purely because of when they teed off. Modeling that requires the weather window for each wave and the difficulty profile of each hole, which is another data dependency layered on top of the skill one.

• Skill splits into components that persist at different rates, so leaderboard history is not enough.
• Field sizes are large, so thin data anywhere in the field distorts the baseline everywhere in it.
• The cut creates a structural break that is correlated with performance rather than independent of it.
• Tee waves mean two players in one event can face different conditions, so the environment is per player.

What shipping anyway would actually cost

A simulation will always produce a number. Give it weak inputs and it produces a number with the same decimal places, the same confident presentation, and none of the meaning. That is the specific hazard of publishing a model without the data behind it: the output looks identical to a well supported one. There is no visual signal on a board that distinguishes a projection built from measured skill components from a projection built from vibes and finishing positions.

It is also not a self correcting mistake in any useful time frame. Golf props settle slowly and noisily. A model that is quietly wrong about how much putting persists could run a full season before its graded record made that clear, and in the meantime it would have published thousands of confident probabilities. Compare that to the sports where we do have data: our commitment there is to publish the whole ledger including the losses, as why we publish our misses lays out. That commitment only means something if the ledger has enough behind it to be worth reading.

What it would take to open the board

The requirements are unglamorous and specific. We need shot level or component level skill measurement for a full field, historically deep enough to fit and shrink player estimates rather than take single season numbers at face value. We need clean tournament results in the same identity space so grading can settle against what actually happened. We need course level detail sufficient to build a hole by hole environment, and per wave conditions to handle the tee time problem. And all of it has to be licensed for commercial use, delivered reliably enough to run on a schedule, and stable enough that a provider change does not silently reshape the model.

Then, before anything ships, the engine has to be backtested on that data the way every other engine was: fit on history, evaluated on periods it never saw, checked for whether its stated probabilities match observed frequencies rather than whether it looks profitable in the sample it was tuned on. Backtesting player prop models covers why that ordering is not optional and why a backtest run on the same window used for fitting is a description of the past rather than a test of anything.

We are not going to put a date on it, and we are deliberately not naming who we are talking to or what anything costs, because both would be a promise dressed up as an update. When the data is in hand and the backtest survives, a golf board will open, and it will open with a graded record from day one like the others.

A coverage map is a claim about everything on it

Here is the part that matters beyond golf. When a research tool covers everything, one of two things is true: either it has solved the data problem in every one of those sports, which is expensive and rare, or its standard for shipping is low enough that the data problem never blocked anything. From the outside those two look the same. The blank spaces are the only visible evidence about which one you are dealing with.

So a gap is information. A product willing to say it does not cover something has demonstrated that it has a threshold, and a threshold applied consistently is what makes the covered sports credible. A product with no gaps at all has told you nothing about its standards, and you are left inferring them from marketing copy. Comparing research tools works through the broader set of questions worth asking; this article is us answering one of them about ourselves before anybody asked.

The same logic runs inside the sports we do cover. Newer engines carry capped grades until their own graded history earns more. Individual offers get blocked when the guardrails detect a situation the model handles badly. Whole markets get suppressed when their measured record falls below a threshold. Every one of those is the same decision as the golf one at a smaller scale: decline to publish confidence that has nothing behind it.

If golf is what you research, we would rather you use something built for it than use a board we shipped to look complete. When ours opens, it will be because the inputs exist and the backtest held, not because the coverage map looked better with one more logo on it. The sports we do run live, and the public record for each of them, are in the Model Room.