Is a draw ever “due”?
No. Neither a long run without a draw (N), nor the draw debt (E), nor the site's own draw score raises the chance that the run ends in the next match or within the next 5. A model that knows nothing about the run predicts draws just as well at N = 0 as at N = 30. The only thing that moves the number is who is playing.
Computed 20 Nov 2026 from the draw model fitted 20 Nov 2026 on 548,141 matches · descriptive statistics, not betting advice
The raw draw rate falls as the run gets longer — 25.6% after a draw, 18.1% after 31+ matches without one: the opposite of “due”. And the model, which has no input about the run at all, draws the same line. The run adds nothing beyond who is playing.
Table
| N | Team-matches | Actual | Model | Difference | 95% ± |
|---|---|---|---|---|---|
| 0 | 206,425 | 25.63% | 25.68% | −0.05 | 0.19 |
| 1 | 153,557 | 25.72% | 25.52% | +0.20 | 0.22 |
| 2 | 114,036 | 25.66% | 25.34% | +0.32 | 0.25 |
| 3 | 84,739 | 25.65% | 25.21% | +0.45 | 0.29 |
| 4 | 62,988 | 25.54% | 24.98% | +0.56 | 0.34 |
| 5 | 46,904 | 25.69% | 24.81% | +0.88 | 0.39 |
| 6 | 34,861 | 24.85% | 24.65% | +0.20 | 0.45 |
| 7 | 26,199 | 24.75% | 24.44% | +0.30 | 0.52 |
| 8 | 19,733 | 24.89% | 24.28% | +0.61 | 0.59 |
| 9 | 14,823 | 24.91% | 24.11% | +0.81 | 0.68 |
| 10 | 11,125 | 23.81% | 23.94% | −0.13 | 0.79 |
| 11 | 8,465 | 24.64% | 23.70% | +0.94 | 0.90 |
| 12 | 6,381 | 23.54% | 23.56% | −0.02 | 1.03 |
| 13–15 | 11,484 | 23.27% | 23.16% | +0.11 | 0.76 |
| 16–20 | 7,007 | 22.66% | 22.38% | +0.29 | 0.97 |
| 21–30 | 2,663 | 20.43% | 20.94% | −0.51 | 1.53 |
| 31+ | 315 | 18.10% | 17.77% | +0.32 | 4.16 |
E is the site's own predictor figure. The draw chance sits flat from the smallest debt to the largest (25.4% in the first bucket, 24.2% in the last), and the actual rate follows the model. Debt does not accumulate probability.
Table
| E | Team-matches | Actual | Model | Difference | 95% ± |
|---|---|---|---|---|---|
| 0–0.5 | 415,240 | 25.44% | 25.32% | +0.12 | 0.13 |
| 0.5–1 | 182,706 | 25.55% | 25.13% | +0.41 | 0.20 |
| 1–1.5 | 100,154 | 25.32% | 24.95% | +0.37 | 0.27 |
| 1.5–2 | 54,860 | 25.46% | 24.87% | +0.59 | 0.36 |
| 2–3 | 44,421 | 25.11% | 24.75% | +0.36 | 0.40 |
| 3–4 | 11,443 | 25.08% | 24.60% | +0.48 | 0.78 |
| 4–6 | 2,790 | 25.41% | 24.52% | +0.89 | 1.58 |
| 6+ | 91 | 24.18% | 26.07% | −1.90 | 8.99 |
Team-matches in groups of equal size, from the lowest score to the highest. The group with the highest score draws 25.0% of the time; the lowest, 23.4%. The score does not rank teams by their chance of a draw.
Table
| Score | Team-matches | Actual | Model | Difference | 95% ± |
|---|---|---|---|---|---|
| 0.004–0.075 | 81,171 | 23.39% | 23.55% | −0.16 | 0.29 |
| 0.075–0.089 | 81,170 | 25.73% | 25.73% | +0.00 | 0.30 |
| 0.089–0.109 | 81,171 | 25.94% | 25.85% | +0.09 | 0.30 |
| 0.109–0.131 | 81,170 | 25.51% | 25.26% | +0.25 | 0.30 |
| 0.131–0.157 | 81,171 | 25.57% | 25.47% | +0.10 | 0.30 |
| 0.157–0.189 | 81,170 | 25.97% | 25.40% | +0.58 | 0.30 |
| 0.189–0.230 | 81,171 | 25.72% | 25.28% | +0.44 | 0.30 |
| 0.230–0.289 | 81,170 | 25.72% | 25.24% | +0.49 | 0.30 |
| 0.289–0.390 | 81,171 | 25.69% | 25.08% | +0.61 | 0.30 |
| 0.390–0.922 | 81,170 | 25.01% | 24.72% | +0.29 | 0.29 |
The three charts above share one scale on purpose, so that flat and falling look like what they are.
This is the question behind a staking line: will it close soon? The longer the run, the less likely it is to end within 5 matches — 77.3% at N = 0 and 70.8% at N = 16–20 — and the actual rate follows the model. The overall offset of +0.6 points is much the same in almost every bucket and does not grow with N: it is a level difference, not a trend. The bucket furthest from the model is N = 31+ (+2.0 points over 310 team-matches): one outlier among about 52 cells on this page, and the windows overlap, so its true error is larger than the whisker shows.
Table
| N | Team-matches | Actual | Model | Difference | 95% ± |
|---|---|---|---|---|---|
| 0 | 204,927 | 77.30% | 76.88% | +0.42 | 0.18 |
| 1 | 152,368 | 77.31% | 76.60% | +0.71 | 0.21 |
| 2 | 113,147 | 77.06% | 76.33% | +0.73 | 0.24 |
| 3 | 84,058 | 76.77% | 76.05% | +0.72 | 0.29 |
| 4 | 62,473 | 76.54% | 75.75% | +0.79 | 0.33 |
| 5 | 46,536 | 76.35% | 75.49% | +0.86 | 0.39 |
| 6 | 34,558 | 75.76% | 75.18% | +0.58 | 0.45 |
| 7 | 25,958 | 75.69% | 74.90% | +0.78 | 0.52 |
| 8 | 19,528 | 75.28% | 74.63% | +0.65 | 0.60 |
| 9 | 14,646 | 74.66% | 74.28% | +0.37 | 0.70 |
| 10 | 11,001 | 74.42% | 74.00% | +0.42 | 0.81 |
| 11 | 8,373 | 74.04% | 73.62% | +0.41 | 0.93 |
| 12 | 6,308 | 73.24% | 73.32% | −0.08 | 1.08 |
| 13–15 | 11,367 | 73.05% | 72.66% | +0.40 | 0.81 |
| 16–20 | 6,890 | 70.80% | 71.13% | −0.33 | 1.05 |
| 21–30 | 2,622 | 68.34% | 68.39% | −0.04 | 1.74 |
| 31+ | 310 | 63.87% | 61.90% | +1.97 | 5.26 |
For every team in every competition season: how many more or fewer draws it had than the model expected, in standard deviations (z). If draws clumped together, the bars would be wider than the curve; if a shortfall were “repaid” later in the season, narrower. They follow the curve: 4.94% of seasons sit outside ±1.96, against 5% for pure chance, and the dispersion ratio — Σ(draws − expected)² ÷ Σ variance — is 0.994.
Table
| z | Team-seasons | Pure chance |
|---|---|---|
| z below -3.5 | 0 | 4 |
| z -3.5 to -3.0 | 10 | 21 |
| z -3.0 to -2.5 | 38 | 91 |
| z -2.5 to -2.0 | 279 | 311 |
| z -2.0 to -1.5 | 802 | 827 |
| z -1.5 to -1.0 | 1,798 | 1,725 |
| z -1.0 to -0.5 | 2,975 | 2,814 |
| z -0.5 to 0.0 | 3,645 | 3,595 |
| z 0.0 to 0.5 | 3,489 | 3,595 |
| z 0.5 to 1.0 | 2,674 | 2,814 |
| z 1.0 to 1.5 | 1,645 | 1,725 |
| z 1.5 to 2.0 | 912 | 827 |
| z 2.0 to 2.5 | 377 | 311 |
| z 2.5 to 3.0 | 93 | 91 |
| z 3.0 to 3.5 | 38 | 21 |
| z 3.5 to 4.0 | 1 | 4 |
| z from 4.0 up | 1 | 1 |
The peak's slight lean to the left is the skew of a count at p ≈ 0.25 (a season cannot have fewer than no draws, but can have many more), not a sign of anything. The totals agree: actual draws are 1.0017 of what the model expected.
- Every played match counts twice, once for each team. N, E, the run against the team's record and the draw score are rebuilt as they stood before the match, with the site's own formulas (config/draw.php, DrawScoreCalculator). A draw is a ninety-minute draw.
- The N, E and score charts use only matches where the team already had 100 or more matches (so p100 exists) and at least one draw (so the run has a known start). A run that began before a team's history does is left out: its length is unknown. 50,445 team-matches fall before a team's first recorded draw.
- The draw score is relative to the whole field, and the field's historical maxima are not stored anywhere, so every past score is normalised against today's field (team_stats as of 20 Nov 2026: largest debt 4.97, longest current run 35).
- “The model” is the site's logistic draw model. Its inputs are the goals model's draw chance, the draw rate at this Elo gap, the expected total goals and the competition's base draw rate — nothing about the run or the debt. It was fitted on these same matches (in sample), but it cannot have learned a run effect, because it never saw one.
- Unlike the records pages, runs here are not broken at gaps in coverage; a gap affects the actual and the modelled side of every comparison in the same way.
- Whiskers are 95% intervals from the model's own spread. There are about 52 cells on this page, so one or two about 2σ from the model are expected noise. The five-match windows of neighbouring team-matches overlap, so their true errors are larger than shown.
- Descriptive statistics over historical results, not betting advice.
Without the 100-match floor the reading is the same: 25.0% after a draw, 18.4% at N = 31+, and the model falls with it. The teams that reach long runs are the teams, leagues and matchups that draw less; given the matchup, the run adds nothing. One wrinkle is statistically real: right after a draw (N = 0) there are 0.25 points fewer draws than the model says (z = −3.00). It is tiny, most likely a calibration artefact, and not a usable effect. With 17 buckets, one at about |z| = 2 is what chance produces.
| N | Team-matches | Actual | Model | z | Clubs actual / model | National teams actual / model |
|---|---|---|---|---|---|---|
| 0 | 264,765 | 25.00% | 25.25% | −3.00 | 25.1% / 25.5% | 24.3% / 22.8% |
| 1 | 196,870 | 25.19% | 25.05% | +1.46 | 25.3% / 25.3% | 23.7% / 22.3% |
| 2 | 146,119 | 25.10% | 24.83% | +2.42 | 25.3% / 25.1% | 23.5% / 22.1% |
| 3 | 108,610 | 24.98% | 24.65% | +2.57 | 25.1% / 24.9% | 23.8% / 21.9% |
| 4 | 80,889 | 24.73% | 24.40% | +2.22 | 25.0% / 24.7% | 22.0% / 21.5% |
| 5 | 60,476 | 24.97% | 24.19% | +4.50 | 25.2% / 24.5% | 22.8% / 21.3% |
| 6 | 45,072 | 24.09% | 23.99% | +0.47 | 24.3% / 24.3% | 21.9% / 21.2% |
| 7 | 33,993 | 24.11% | 23.74% | +1.65 | 24.4% / 24.1% | 21.3% / 20.7% |
| 8 | 25,633 | 23.65% | 23.54% | +0.40 | 23.9% / 23.9% | 21.6% / 20.8% |
| 9 | 19,447 | 24.20% | 23.33% | +2.89 | 24.5% / 23.7% | 22.0% / 20.6% |
| 10 | 14,622 | 22.78% | 23.11% | −0.95 | 23.0% / 23.5% | 21.3% / 20.1% |
| 11 | 11,206 | 23.42% | 22.87% | +1.40 | 24.0% / 23.3% | 19.1% / 19.8% |
| 12 | 8,519 | 22.90% | 22.63% | +0.61 | 23.3% / 23.0% | 19.8% / 19.7% |
| 13–15 | 15,455 | 22.21% | 22.19% | +0.07 | 22.6% / 22.6% | 19.6% / 19.4% |
| 16–20 | 9,781 | 21.24% | 21.31% | −0.19 | 21.7% / 21.8% | 18.7% / 18.6% |
| 21–30 | 3,919 | 19.55% | 20.00% | −0.72 | 20.2% / 20.6% | 16.7% / 17.6% |
| 31+ | 468 | 18.38% | 16.95% | +0.84 | 21.5% / 18.2% | 13.4% / 15.0% |
And by how close the run is to the team's own record (N ÷ the longest run so far) — the predictor's “stretch”. At the record, the draw rate is 24.7% against 24.1% from the model.
| Run ÷ record | Team-matches | Next match actual / model | Within 5 actual / model |
|---|---|---|---|
| under 25% | 606,919 | 25.45% / 25.26% | 76.81% / 76.18% |
| 25–50% | 136,848 | 25.54% / 24.96% | 76.46% / 75.68% |
| 50–75% | 44,219 | 25.18% / 24.81% | 75.50% / 75.43% |
| 75–100% | 13,397 | 24.57% / 24.53% | 75.22% / 74.90% |
| at the record | 10,322 | 24.66% / 24.14% | 74.76% / 74.23% |
Test 18,777 team-seasons and about 5% of them look “unusual” on chance alone. After the Benjamini–Hochberg adjustment, which asks what share of the rows called unusual would be noise, 0 team-seasons and 0 of 3,044 league-seasons stay below q = 0.05. Those are the rare genuinely unusual cases; everything else in these tables is what chance produces. Whole league-seasons do scatter more than independent matches allow (dispersion ratio 1.182): some competition seasons draw more or less than the model's all-history competition rate. That is a season's climate, shared by every team in it — not a team repaying a debt. Tails use the exact Poisson-binomial distribution (a skew-corrected normal above 1,000 matches).
| Team-seasons | Matches | Drew | Expected | z | q (BH) |
|---|---|---|---|---|---|
| Rio Ave, Primeira Liga 2023/24 | 34 | 19 | 8.6 | +4.13 | 1.000 |
| Chicago Fire, MLS 2014 | 34 | 18 | 8.9 | +3.55 | 1.000 |
| Stranraer, Scottish League One 1999/00 | 36 | 18 | 9.0 | +3.47 | 1.000 |
| Basel, Super League 2021/22 | 36 | 17 | 8.3 | +3.44 | 1.000 |
| 1. FC Kaiserslautern, Bundesliga 1995/96 | 34 | 18 | 9.2 | +3.41 | 1.000 |
| Germinal Beerschot, Belgian Pro League 2001/02 | 34 | 16 | 7.8 | +3.38 | 1.000 |
| CD Badajoz, Segunda División 1999/00 | 42 | 24 | 13.7 | +3.38 | 1.000 |
| Nashville SC, MLS 2021 | 36 | 19 | 9.9 | +3.38 | 1.000 |
| SSC Napoli, Serie B 2003/04 | 46 | 26 | 15.3 | +3.35 | 1.000 |
| Southend United, League Two 2002/03 | 46 | 3 | 13.3 | −3.34 | 1.000 |
| League-seasons | Matches | Drew | Expected | z | q (BH) |
|---|---|---|---|---|---|
| League Cup League Cup 2017/18 | 93 | 5 | 20.0 | −3.82 | 0.098 |
| Club Friendlies 2024 | 152 | 51 | 31.5 | +3.93 | 0.186 |
| Bundesliga 1995/96 | 306 | 108 | 79.3 | +3.76 | 0.186 |
| League Cup League Cup 2014/15 | 93 | 7 | 21.1 | −3.50 | 0.186 |
| FA Cup FA Cup 1887/88 | 144 | 10 | 25.7 | −3.43 | 0.186 |
| UEFA Europa League UEFA Europa League 1972/73 | 124 | 12 | 27.3 | −3.32 | 0.215 |
| Scottish League Two 2006/07 | 180 | 22 | 40.4 | −3.30 | 0.215 |
| League Cup League Cup 2016/17 | 93 | 8 | 20.8 | −3.20 | 0.215 |
| League Cup League Cup 2013/14 | 93 | 8 | 20.8 | −3.19 | 0.215 |
| Belgian Cup 1934/35 | 69 | 1 | 9.6 | −2.99 | 0.215 |
For every draw, did the next k matches all avoid one? The expected count is what independent matches with the model's own chances produce: summed over the same draws, the product of each following match's chance of not drawing. Up to runs of 30 the counts agree within 10%: the runs on this site are the length chance produces, given who played. At 50+ there are 2 against 2.2 expected — in line with independence.
| Run of | Draws followed by k priced matches | Runs observed | Expected | Observed ÷ expected | Clubs | National teams |
|---|---|---|---|---|---|---|
| 5+ | 258,014 | 60,863 | 61,488 | 0.99 | 55,031 / 55,189.8 | 5,832 / 6,298.2 |
| 10+ | 252,283 | 14,726 | 15,187 | 0.97 | 13,073 / 13,353.7 | 1,653 / 1,833.2 |
| 15+ | 247,731 | 3,887 | 4,001 | 0.97 | 3,371 / 3,447.2 | 516 / 553.9 |
| 20+ | 244,039 | 1,106 | 1,127 | 0.98 | 932 / 953.0 | 174 / 173.9 |
| 25+ | 240,424 | 364 | 338 | 1.08 | 292 / 281.4 | 72 / 56.9 |
| 30+ | 237,089 | 114 | 109 | 1.05 | 87 / 89.0 | 27 / 19.7 |
| 40+ | 231,220 | 12 | 13.6 | 0.88 | 7 / 10.8 | 5 / 2.8 |
| 50+ | 226,075 | 2 | 2.2 | 0.93 | 1 / 1.6 | 1 / 0.5 |
The expectation conditions on where the draws actually fell: under independence a draw says nothing about the matches after it, so this has the same expected value as the fully modelled count and needs no model of where draws land. A run still going counts once it has reached k; a draw too close to the end of a history to have k matches after it is in neither column. The rows are cumulative, so they are not independent of each other.
Computed 20 Nov 2026 by football:index:draw-due from the draw model fitted 20 Nov 2026. Related: luck or skill · the draw predictor · not betting advice.