Curling statistics and analytics
Curling broadcasts are full of percentages that nobody defines. Here is what each one measures, what a good one looks like, and which of them are actually worth anything.
Last updated 2026-08-18
Shot percentage, and what is wrong with it#
The number every broadcast leads with. Its definition is simple: an official statistician scores every stone out of four.
| Score | What it means |
|---|---|
| 4 | The shot was made as called, completely. |
| 3 | Most of what was wanted. A draw that is a foot heavy but still counting. |
| 2 | Half of it. The stone is in play and does something, but not the thing. |
| 1 | Almost nothing. Contact made, or the stone in play, and no more. |
| 0 | Complete miss. Through the house, hogged, wrong stone removed. |
A player's percentage is their points divided by the maximum they could have scored. Eighty-eight per cent means averaging about three and a half out of four. It is intuitive, it is quick to collect, and it has three problems the sport has known about for decades.
It is blind to difficulty. A lead throwing two guards to an
empty sheet and a skip throwing a double for four are both marked out of
four. Leads therefore shoot higher than skips, always, everywhere, and it
tells you nothing about who played better.
It rewards the safe call. A skip who takes the easy hit
instead of the hard freeze protects their percentage and possibly loses the
end.
It is scored by a person. "Most of what was wanted" is a
judgement, made quickly, by somebody who was not in the huddle and does not
know what was actually called.
None of which makes it useless. Within one position, across a week, it is a perfectly reasonable measure of execution: a skip shooting 78% has genuinely thrown worse than a skip shooting 89%. What it cannot support is comparison between positions, or between teams playing different styles — and those are exactly the comparisons it gets used for.
What counts as a good number#
Percentages mean nothing without a scale, and the scale is rarely given.
- ~85%team average at a world championship
- ~80%a strong competitive club team
- 4points available per stone
- 16stones graded every end, eight per team
The compression at the top is the striking part. Elite curling is separated by three or four percentage points, which at sixteen stones an end over ten ends is a handful of stones across a whole game. This is why the sport is decided by which shots are missed rather than how many.
The metrics that measure a team#
These are better. Each measures whether a team is achieving the thing the situation calls for, rather than whether individual stones went where they were pointed.
| Metric | What it measures | What good looks like |
|---|---|---|
| Hammer efficiency | Of the ends a team scored in with the hammer, the share in which it took two or more. The hammer is meant to produce a multi-point end; this measures whether it did. | 45% or better indicates a very good team. |
| Force efficiency | How often a team without the hammer holds its opponent to a single point. Failing to force means giving up two or more, which is how games get away. | Higher is better; it is the defensive twin of hammer efficiency. |
| Steal efficiency | Ends stolen — scored without the hammer — as a share of all ends played without it, blanks included. | Stealing is the single most valuable outcome in curling, and rare. |
| Steal defence | How often a team gives up a steal when it has the hammer. The mirror of steal efficiency, and a good marker of composure. | Lower is better. |
Hammer efficiency has the most interesting provenance. It comes from an idea proposed by the Canadian champion and Olympian Linda Moore: that what really distinguishes teams is not how many shots they make but how often they convert last stone into a genuine advantage. That has held up better than most curling analytics of its era.
What the hammer is worth#
Roughly three quarters of a point. That single number explains an enormous amount of what teams do, and it is worth sitting with.
If last stone is worth about 0.75 points, then scoring exactly one point with it is close to a break-even outcome and arguably a loss: you gained one point and surrendered an asset worth three quarters of one. Whereas a blank end costs you nothing and keeps the asset. And a steal — scoring without the hammer — gains you the point and keeps the hammer, which is why it swings games so hard.
Every row below is written from your team's point of view. The rule driving the third column is the only one that matters: whoever scores throws first in the next end, and throwing first means not having the hammer.
| What happened to your team | Points | Who has the hammer next end | Roughly worth |
|---|---|---|---|
| You steal — scored without the hammer | +1 or more | Your opponent, again | Enormous. Points from an end you were expected to lose. It does not win you last stone, which is why stealing repeatedly is so hard. |
| You score two or more with the hammer | +2 or more | Your opponent | The target outcome for a hammer end, and what hammer efficiency measures. |
| You blank with the hammer | 0 | You keep it | Neutral to slightly positive, and often better than taking one. |
| You score one with the hammer | +1 | Your opponent | Close to neutral. You gained a point and gave up an advantage worth about three quarters of one. |
| You force your opponent to one — they had the hammer | 0 | You | A win for you. They spent the hammer on a single point and it now comes to you. |
| You give up a steal — you had the hammer and did not score | 0 | You keep it | The worst common outcome. You still have last stone, which is the only consolation. |
Where the analytics disagree with the folklore#
Curling has a real analytics literature now, and it is worth knowing where it contradicts what commentators say.
Blank rather than take one#
The most robust published finding. A three-dimensional Markov model built on eighteen years of Canadian championship data concluded that a team with hammer faced with the choice between one point and a blank should usually blank. Good skips already did this; the value of the work is that it turns a habit into a quantity, and identifies the score-and-end situations where the habit is wrong.
Machines pick better hammer shots than Olympians#
A 2016 paper presented a search-based shot selector, tested against real hammer situations from the 2010 Olympic Games, and reported a statistically significant improvement over the outcomes the Olympic teams themselves achieved. Take it with the usual caveat — a simulator cannot miss the way a human misses, so "best shot" and "best shot for this team on this ice tonight" are different questions. It is the same caveat that applies to any engine playing this game, ours included.
Expected score beats shot percentage#
The direction of travel is away from grading individual stones and toward modelling the whole board: given this arrangement of stones, this end, this score, what is the distribution of outcomes, and how much did that shot move it? That is the right question, and it is the one a shot percentage cannot ask. Recent work builds these models on real championship board positions rather than on simulated ones.
The numbers on this site#
For what it is worth, this site takes a position on the metric problem by measuring something different. Rather than grading a shot against what was called — which requires knowing what was called — the analysis here compares the board a shot actually left against the boards the alternatives would have left, using the same search the computer opponent uses. That is closer to expected score than to shot percentage, and it has the practical advantage of not needing a statistician in the stands.
It is also, unavoidably, an opinion about which shot was best. Every analysis tool in this sport is.
Elsewhere#
- CurlingZoneThe statistical record of competitive curling: results, rankings, shot percentages, team histories. Where most curling analytics starts.
- Throwing Rocks — Glenn PaulleyA statistician writing seriously about curling metrics, including the scoring metrics broadcasters use without explaining them.
Sources#
- Primary An analysis of curling using a three-dimensional Markov model Journal of Sports Analytics, 2019The win-probability model and the blanking conclusion.
- Primary Action selection for hammer shots in curling IJCAI, 2016The shot selector tested against 2010 Olympic hammer situations.
- The evolution of curling analytics MIT Sloan Sports Analytics ConferenceHow the sport moved from shot percentages to expected-score models, and the origin of hammer efficiency in Linda Moore’s proposal.
- Building an all-shot expected-score distribution model from real-match curling boards Applied Sciences, 2026The current state of expected-score modelling, trained on real championship boards.
The decisions these numbers are about The rest of the literature