The Model Moves the Market

Why this study exists

Our own model page carries an open question we have never been able to answer: would following the model have made money? Accuracy and profit are different things. A model can rank every player correctly and still lose, because the price already contains most of what it knows and the book keeps the difference.

We have now archived seven seasons of sportsbook prices — opening and closing, per player, per market, per book, already graded against what happened. That makes the question answerable rather than arguable.

01 The model predicts where the line goes

Closing-line value is the standard test of whether a bettor has an edge. If you consistently take prices that shorten before the market closes, you are seeing something ahead of the people setting them. It is a cleaner measurement than profit because it does not have to wait for variance to average out.

We took every outcome where the model’s probability implied a positive expected value at the best opening price available, and measured what happened to that price.

Every market. Every threshold we tested. And it is monotone in confidence: raise the bar for what counts as a bet and the closing-line value goes up rather than down. On top-10 outcomes it climbs from +21% at any positive edge to +84% at the strictest cut. A model that was picking at random would show no such gradient.

02 Whether it makes money is unproven, in both directions

Here is the same set of bets, settled. One unit on each, priced at the open, across 138 tournaments the calibration was never fitted on.

Some individual cells look excellent. Applying the calibration correction and betting only outcomes with a 20% modelled edge returned +46.6% on top-5 markets. That number is worthless: it rests on 145 bets at an average price of 77, where a single winner swings the result by more than fifty units, and its confidence interval runs from −100% to +342%.

The stronger variant, shown rather than described

DataGolf archives two model variants. The one that folds in course history does noticeably better, and reporting only the weaker of the two would be its own kind of overclaim.

Read the intervals rather than the point estimates. A strategy returning +33% whose plausible range runs from −78% to +203% has told you almost nothing, and 305 bets is nowhere near enough to narrow it. That is the honest position: the quantity we can measure precisely here is the line movement, not the profit.

03 The hold is the whole story

These findings are not in tension. The model sees something real — that is what the closing-line value proves. What is unresolved is whether it sees enough of it to pay the toll.

Set the two numbers side by side. The model’s prices shorten by around 20% of what was taken. The hold on those same markets runs from 16% to 44%. The edge is real and it is about the size of the fee — which is exactly the situation in which you reliably move the line and cannot tell, across seven seasons, whether you are ahead of it.

This also explains the one book that behaves differently. Pinnacle charges a third of what the recreational books charge on outright winners, which is the same finding our third study reached from the opposite direction.

04 Fixing the probabilities does not fix the profit

The model’s raw probabilities are miscalibrated in a consistent direction, and that is correctable. Fitting a two-parameter correction on 2019–2022 and applying it to seasons the fit had never seen cuts calibration error sharply.

The correction improves closing-line value too — in all four markets, at every threshold. It makes the model demonstrably better at the thing it is for.

It does not settle the profit question. Corrected returns bounce between +46% and −42% depending on which cell you look at, with intervals wide enough to contain both. Better probabilities help; they do not narrow the interval enough to resolve it.

What we are not claiming

• That betting golf cannot be profitable. We tested one public model, four outright markets and three books. Matchups, live markets and a private model are all outside this.
• That the opening price was really available. We took the best opening number across three books. Limits, line moves and closed accounts are not in the archive, and every one of them makes the real result worse than the one shown here.
• That closing-line value is worthless. It is the most robust signal in this study. It proves the model carries information. It does not prove that information survives the price.

The one thing we will claim is the negative: we have looked, with seven seasons and 120,065 graded outcomes, and we cannot find a strategy off these models whose profit is distinguishable from luck — in either direction. Several look profitable. None survives an honest confidence interval. Anyone selling you an edge off the same public inputs owes you this table, and so do we.

Method

• Model probabilities read from an archive of pre-tournament predictions, stored as they stood before each event. Recomputing them from today’s ratings would leak the outcome into the input.
• Bets placed at the best opening price across the three books. Closing prices used only to measure line movement.
• Calibration fitted on 2019–2022 and evaluated only on 2023–2025.
• Both archived model variants tested — the plain baseline and the one that folds in course history. The second performs better and reaches the same verdict.
• Flat one-unit stakes. No Kelly sizing, which would change the variance but not the sign.
• Confidence intervals from 4,000 bootstrap resamples drawn at tournament level.
• Outcomes were already graded in the source data; nothing was scored by us.