Can You Actually Predict Head-to-Head Golf? We Tested 220,000 Matchups.

At Elo Golf, we rank the world's golfers — from PGA Tour pros to your average weekend player — using the Elo rating system, the same algorithm that powers chess rankings. But Elo doesn't just rank players. It also predicts who's more likely to win when two or three golfers are paired together in a tee time group.

So we put those predictions to the test. Over 220,000 head-to-head matchups across the PGA Tour — that's every 2-ball and 3-ball pairing within a tee time group, predicting who posts the lower score each round. Here's what we found.

First, how do you measure a prediction?

Before we dive into the details, it's worth explaining what we're measuring. A prediction model isn't accurate just because it picks the winner more than half the time. What matters is calibration.

Here's the test: if our model says a player has a 65% chance of beating their opponent, and we look at every matchup where we said 65%, do those players actually win 65% of the time? If so, the model is well-calibrated. If they only win 55% of the time, the model is overconfident. If they win 75% of the time, it's underconfident.

This idea isn't new — it's how meteorologists evaluate weather forecasts, how FiveThirtyEight graded their election models, and how FIDE evaluates the chess Elo system that inspired our approach. We're applying the same standard to golf.

The results

In the chart below, the dashed grey line represents the perfect calibration — if we predict 60%, the goal is the player wins 60% of the time. The green line shows how our model actually performs.

The green line tracks the diagonal grey line closely. Sure, it's not perfect, but no prediction algorithm is. Looking at the data you'll see the green line sits slightly above the grey across most of the range. That means our model is very slightly underconfident. When we say 57.2%, the actual win rate is about 57.8%. When we say 62.1%, it's about 62.6%.

Across all prediction buckets, the average error is about 1.5 percentage points. For a sport where a top-10 player can lose to a player ranked 150th on any given day, that's pretty strong. You'll also notice slightly larger gaps at the edges of the chart — the 72% bucket is off by about 4.5%. That's not a model problem — it's a data problem. The blue bars in the chart show the sample size for each bucket — the middle buckets (around 50/50 matchups) have 50,000+ data points each, while the extreme buckets have under 1,000. With smaller samples, the margin of error is naturally wider. As more PGA Tour rounds are played, those edge buckets will tighten up too.

For the stats nerds: we measured this formally using the Brier score, the standard metric for evaluating probabilistic predictions. Our model achieves a reliability score of 0.00023 — for context, anything below 0.001 is considered excellent calibration.

Why aren't there any 80%+ predictions?

You might have noticed the chart only ranges from about 28% to 72%. Where are the 80% or 90% predictions? In chess, a grandmaster might have a 95% chance of beating a club player. Why doesn't golf work the same way?

Part of the answer is that this data only includes PGA Tour players — the best golfers in the world. Even the "weakest" player in a PGA Tour field is extraordinarily good at golf. If we were comparing a tour pro against a club amateur, you'd absolutely see 90%+ predictions — and as we collect more amateur data, that's exactly the kind of cross-skill prediction we'll be able to make. But within the tour, the skill gaps are relatively narrow.

The other part is that golf is one of the most random sports on the planet. A top-10 player can shoot 75 one day and 65 the next on the same course. There's no home-court advantage. Weather changes between tee times. And unlike chess or tennis, you can't play cautiously when you're ahead — the course doesn't care about your rating.

A naive application of the Elo rating system would produce 80%+ predictions for large rating gaps. But when you check those predictions against reality, they're wildly overconfident — the favorites don't win nearly as often as the math suggests. This is a well-documented phenomenon across high-variance sports, from FIFA's football rankings to FiveThirtyEight's NBA model.

Our approach separates two things that most Elo systems conflate: the formula that updates ratings after a match, and the formula that converts ratings into predictions. The update formula stays fixed (so ratings remain comparable over time), but the prediction formula is tuned to match golf's actual level of randomness. This is the same approach recommended by Glicko and TrueSkill researchers. The result is a narrower, more honest range of predictions — because that's what the data actually supports.

What about draws?

One important caveat: golfers tie more often than you'd think. In our dataset, about 10% of head-to-head matchups end in a draw — both players post the same score for the round. The Elo system doesn't predict the probability of a draw separately. It predicts "Player A beats Player B" and treats a tie as half a win.

This means that when a strong favorite ties with an underdog, the favorite loses Elo points — because a tie is worse than the expected outcome for a player who was supposed to win. That's by design. Over time, if a player keeps tying when they "should" be winning, their rating adjusts downward to reflect reality.

This is similar to sports betting, where a tie is often (but not always) a push — your bet is returned. So our win probability maps cleanly to the question a 2-ball bet is actually asking: "will this player beat the other one?"

What this means for you

If you follow the PGA Tour, our upcoming matches page shows real-time predictions for every tee time group. When Scottie Scheffler is paired with a player ranked 80th in the world, you'll see the actual probability of each player posting the lower score — not a guess, but a number backed by 220,000+ data points tracked since 2012.

If you're into golf betting, calibrated probabilities are what you need to make informed 2-ball and 3-ball bets. A model that says "70%" when the true probability is 60% will cost you money over time. Our predictions give you a more honest starting point.

But here's the best bit. The same math works for any golfer with enough match history. The Elo system doesn't care if you're Scottie Scheffler or a 15-handicap playing your club championship. It just takes your scorecards — amateur or pro — and each round builds a picture of how you perform against not only the people you play with but anyone around the world.

Imagine your club championship with real-time Elo predictions for every match. Maybe you lost your round today, but you can walk off the 18th and confidently tell your playing partner: "I'll get you next time — I win 64% of the time."

Sign up and submit a scorecard to see where you rank locally and globally, and find out how likely you are to win your next tee time battle.

Where to from here?

We re-evaluate our prediction model regularly, checking that our calibration holds as new data comes in. Golf changes — players improve, courses get renovated, equipment evolves — and our model needs to keep up.

We also publish the Elo ratings and predictions openly. You can see every player's rating, their history, and the predictions for upcoming rounds. If you think we're wrong about a matchup, you can check back after the round and see for yourself.

And remember — a 60% favorite still loses 40% of the time. Golf is one of the most fun sports in the world, but it's also one of the most unpredictable. Enjoy your golf, and enjoy the data that comes with it.

Have questions about how Elo Golf works? Check out our how it works page or read about what Elo rating means in golf. Want to see predictions in action? Head to the upcoming matches page.