Why Are Draws So Hard to Predict? Inside World Cup 2026's Draw-Fest

Spain, ranked among the favourites for the whole tournament, walked off after a goalless draw with Cape Verde. Brazil were pegged back to 1-1 by Morocco. Portugal could only manage 1-1 with DR Congo. Across the opening fortnight of World Cup 2026, nine of the first 21 finished matches have ended in a draw — about 43%. That is roughly double football's usual rate: across major European leagues and World Cup history, draws cluster in the low-to-mid 20s — call it one in four, around 25% (The Analyst; WFStats). No league or World Cup era on record has ever *sustained* 40%+ (Engora Data Blog).

Which raises an honest question we get asked a lot: if draws are this common, why does no prediction model — ours included — ever seem to 'call' one? The answer isn't that models are blind to draws. It's that the draw is, mathematically, the single hardest outcome in football to point at with confidence. Here's why.

The modal-outcome trap

In a football result there are three boxes: home win, draw, away win. A model that has to name one result picks whichever box holds the most probability — the 'modal' outcome. The problem is that the draw almost never holds the most. To be the top pick, the draw has to beat *both* the home win and the away win at the same time, and that only happens in a thin sliver of perfectly balanced, low-scoring fixtures.

The scale of the gap is startling. In one study of 12,337 matches, draws actually occurred in 25.9% of games — yet the model rated 'draw' as its single most likely outcome in just 4.3% of them (Foresportia). The model wasn't wrong about draws. It can say 'draw 28%' all day. It just rarely says 'draw is *the most likely* thing here', because a clear favourite can be 60% or 80% to win while the draw is structurally capped near a third (more on that below).

The knife-edge: a draw is a thin diagonal

Picture every possible scoreline as a grid — 0-0, 1-0, 0-1, 2-1, 1-1 and so on. The draws sit on a single diagonal: 0-0, 1-1, 2-2. Everything off that diagonal is a win for one side. There are simply far more off-diagonal scorelines than on it, so the draw lives on a knife-edge: one goal in either direction tips 1-1 into 2-1 or 1-2, and the draw is gone.

This is why bookmaker and model draw probabilities run into a ceiling around 33%. On years of Premier League data the implied draw probability averaged 26.1%, with a maximum of just 33.3% and only a handful of matches ever crossing 31% (Pena.lt/y). The draw probability peaks when two teams are evenly matched and the game is low-scoring — but it physically can't climb much past a third, because there are only so many ways to finish level.

Low scoring turns one fluke into the whole story

Football is starving for goals. Even Qatar 2022, the World Cup with the most goals ever (172), averaged just 2.69 goals a game, and 34 of its 64 matches finished with fewer than three goals (Wikipedia). When goals are this rare, the margin between a draw and a win is a single moment — a deflection, a penalty, a goalkeeping slip, a late corner. Statisticians model goals as a Poisson process precisely because they behave like rare, semi-random events (Tactiq). In a high-scoring sport the better team's edge has dozens of chances to show; in football one fluke can erase ninety minutes of superiority and leave you with a draw nobody could have priced.

The cruel overlap: even games are both likeliest to draw and hardest to call

Here is the trap that makes the draw uniquely awkward. A draw is most probable when two sides are closely matched — but a 50/50 contest is, by definition, the hardest game to call for *any* outcome, because the probability is smeared across all three results with no dominant favourite (The Punters Page). The exact conditions that make a draw likely are the conditions that make the whole result a coin-flip. So even when the draw climbs to its ceiling, the two win probabilities are usually sitting right beside it — and the draw is still only the second-largest of the three.

Game-state incentives that ratings can't see

Once a match is actually level late on, teams stop behaving the way their pre-match ratings imply. Sometimes both sides quietly accept the point — a draw is enough to advance from a group, or to stay up — and they close the game down, locking the draw in. Sometimes one side throws bodies forward chasing a winner, which blows the game open and produces a late goal at either end, destroying the draw. Which way it breaks depends on the live score, the clock and what a point is worth to each team that day — none of which a flat power rating contains. The infamous 1982 'Disgrace of Gijón', where West Germany and Austria stopped competing once a 1-0 suited both, was a textbook game-theory equilibrium; FIFA's fix was to make the final group games kick off simultaneously (Wikipedia). These incentives peak in group openers, where caution is high, a point is acceptable and losing your first game is costly — exactly the conditions we're watching unfold across the groups right now.

The maths itself had to be patched

Here's the tell that draws are genuinely hard: the standard goals model gets them *wrong* until you fix it by hand. The classic approach treats home and away goals as independent Poisson draws — but in 1997, Dixon and Coles showed that this independence assumption under-counts exactly the low scores that produce draws: it underestimates 0-0 and 1-1 and overestimates 1-0 and 0-1 (Grokipedia). Their fix bolts on a correction that nudges probability back onto 0-0 and 1-1 — the draw scorelines. Refits of real data keep finding the same negative dependence (around -0.13) that inflates those level results (dashee87). In plain terms: the naive model under-predicts draws, and serious modellers have known it for nearly thirty years.

What our own numbers show

We publish our misses, so here is the uncomfortable mirror. Across the nine draws so far, our model's pre-match probability *for the draw* ranged from just 14% to 29%. And in zero of the nine — 0/9 — was the draw our single highest-probability outcome. A home or away win was always rated more likely. Spain 0-0 Cape Verde, we had Spain at 83% and the draw at 14%. Netherlands 2-2 Japan, we had it 44/29/28. Brazil 1-1 Morocco was 45/28/27. Every time, the draw was the bridesmaid.

That looks bad if you only count 'did it pick the winner'. On that scoreboard we're 11/21, 52%, and the draws are dragging it down. But hit-rate is the wrong test for draws — and here's the honest distinction. There are two different questions: *did you name the right result?* and *were your probabilities right?* For the second, statisticians use proper scoring rules like the Ranked Probability Score (RPS), where lower is better and you're rewarded for honest probabilities, not lucky guesses. Our RPS is 0.162 — better than the 0.206 a no-information uniform guess would score. So the model is genuinely adding probabilistic value; it just can't convert that into 'picking' draws, because by the maths above the draw is never its top box. That isn't an excuse. It's the whole point: a model can be *good* at draws — well-calibrated, correctly nudging probability onto level scorelines — and still 'predict' almost none of them. You can check the live tally on our record.

How to actually read a draw probability

So when you open a fixture on our matches page and see a draw at 28%, don't read it as 'we expect a draw'. Read it as coin-flip energy — *these teams are genuinely level, and the result is close to a toss-up across all three outcomes.* A high draw percentage is the model's way of saying 'don't trust the favourite here', not 'back the draw'. That distinction is the single most useful thing to take from this piece, and it's why we'd rather show you honest three-way probabilities than a confident single pick that pretends the draw doesn't exist.

We'll keep publishing every probability and every miss — including the ones the draws cost us — because a model you can audit is worth more than one that only shows you its winners. If you want the full machinery, it's all laid out in how the model works.