Weekly review
Week of 7 September
7 September 2026 – 13 September 2026 · 1,549 words
Between 7 September and 13 September we had published a prediction before kick-off for 226 fixtures, across 7 days and 24 competitions. We called 105 of them correctly, 46.5%.
Does our confidence mean anything
Taking every call in the period and sorting it by how confident we were: we said under 40% on 32 fixtures, averaging 38.2%, and were right 43.8% of the time; we said 40 to 50% on 88 fixtures, averaging 44.8%, and were right 36.4% of the time; we said 50 to 60% on 51 fixtures, averaging 54.7%, and were right 41.2% of the time; we said 60 to 70% on 36 fixtures, averaging 63.9%, and were right 63.9% of the time. The band that misses is 50 to 60%: we overstated it by 13.5 points across 51 fixtures. That is the number to hold us to, and it is the one we would most like to see shrink. At the other end, 60 to 70% came out within 0.0 points of what we claimed, on 36 fixtures — there, the figure on the card means what it says.
The outcome we almost never call
Of 226 fixtures, 60 ended level — 26.5% of them. We called a draw in none of them. Not rarely: never, and that is worth explaining rather than apologising for. The three probabilities on each card come from simulating the two sides' scoring rates independently and counting how often each outcome appears. That method has a built-in ceiling on the draw: across this period the highest draw probability we published was 27.7%, and the nearest a draw ever came to leading its fixture was 8.9 points short. For the draw to top a card, two teams would have to be so exactly matched that neither win is likelier than the tie — which is rarer than a tie itself. So the headline call cannot name the outcome that happens a quarter of the time, and roughly a quarter of our misses are lost before kick-off. Independence is also the part we would change first. Real football scores are not independent at the low end: 0-0 and 1-1 happen more often than two separate rates predict, which is why the standard correction in the literature exists and why our draw numbers are likely a little low across the board. Until that is fixed, the probability we print for the draw is the figure worth reading on a tight fixture — not the call above it.
Home, away, and which we get right
We backed the home side 169 times and the visiting side 57, so 75% of our calls went to the home team. They landed 46.7% of the time. The away calls landed 45.6%.
When we disagreed with the bookmakers
On 193 fixtures we had both our own number and a bookmaker price for the outcome we called. We were at least 10 points above the market on 32 of them — a real disagreement, several times larger than the margin built into the odds — and those calls came in 25.0% of the time. On the 67 fixtures where our number sat within 5 points of the price, we were right 47.8%. Read that in the least flattering direction, because it is the correct one: when we part company with the market, the market is usually the one that was right. A disagreement is where a model claims to know better, and over this period ours did not earn the claim. It is a reason to treat our boldest departures as the weakest part of the output rather than the most interesting, and we would rather print that here than let it sit unmeasured.
Against the bookmakers' favourite
The benchmark worth measuring against is not chance, it is the shortest price — anyone can read the odds for nothing. On the 192 fixtures where we had all three prices, backing the bookmakers' favourite every time would have been right 94 times, 49.0%. We were right 89 times, 46.4%. We named the same outcome as the market on 156 of them, 81%. So over this period the model came out 2.6 points BELOW simply following the favourite. That is the least comfortable number on this page and it is the one we would want a reader to see first. It does not make the probabilities worthless — a calibrated number tells you something a binary pick never does, and the market has no obligation to explain itself — but anyone here for the pick alone should know that the free version of it is currently doing better.
The goals we expected
Underneath the win probabilities sit two expected-goals figures, and they are the part of the model that is easiest to check and hardest to fudge. Across 226 fixtures we expected 1.57 goals for the home side and 1.18 for the away side, a total of 2.76. What actually happened was 1.60 and 1.27, a total of 2.87. Both sides of the scoreline land within a tenth of a goal, which is as close as this kind of forecast gets. Rounding each expectation to a whole number reproduced the exact scoreline 23 times, 10.2%. We mention that as a curiosity rather than a claim: nobody should be predicting exact scorelines, and a model that appeared good at it would be a model that had been overfitted.
The high-scoring flag, tested
One model on each card does nothing but estimate how likely the home side is to score four goals or more. Over 226 fixtures that happened 19 times, 8.4%. The flag was raised on 215 of them, 95% of what we published. That is too often to mean anything: a warning attached to most fixtures is a decoration rather than a warning, and we would rather say so than leave it decorating the cards. Where it was raised, the home side reached four goals 8.8% of the time; where it was not, 0.0%. So it is separating the two, by a factor of about 88.4. That is what the flag is asked to do, and worth stating plainly given it is also the part of the output we have most recently had to correct.
The calls we backed hardest
19 calls were published at 70% or above. 15 of them came in, 78.9%, against an average stated confidence of 77.3%. The most expensive of them was AZ Alkmaar to beat Willem II at 86.3%, which finished 1–1 — a draw, which is the outcome we structurally cannot call.
The call we would point to
The call we are most pleased with is not the most confident one, it is the one furthest from the price. RAAL La Louvière to beat KV Kortrijk in Belgian Pro League: we published 67.3% where the odds implied about 42.6%, a gap of 24.7 points, and it finished 3–0. 8 calls in this period were that far above the market and came in. We list one rather than all of them because a page of the ones that worked, with the ones that did not left out, is how a track record gets quietly turned into advertising — the figure for how that whole group did is in the section above, and it is not flattering.
Where we were most wrong
AZ Alkmaar against Willem II, in Eredivisie, finished 1–1. We had published AZ Alkmaar to beat Willem II at 86.3%, the largest confidence we placed on anything that did not happen in this period. It is worth noting what our five worst misses have in common: 3 of them ended level. The model's most confident errors are not upsets, they are draws. It stays on the record. A site that reports only its hits is not reporting a record, it is reporting a selection.
Against last week
Against the week before: 240 settled fixtures then at 48.8%, 226 now at 46.5%, a move of -2.3 points. More useful than the headline: in the 50 to 70% range, where most of our calls sit, we were 10.3 points below our stated confidence last week and 7.9 points below it this one. That gap narrowing is the only kind of improvement we would claim. The headline difference itself is small enough that we would not call it a change at all.
Where we published
We published across 24 competitions. The five busiest were Championship (19), League One (13), UEFA Champions League (12), League Two (12), Liga Profesional Argentina (11). None of them has reached 20 settled fixtures in this period, which is the least we would want before ranking one competition against another.
How to read all this
One week of 226 fixtures settles very little on its own. A percentage point either way on a sample this size is noise, and we would rather say so than dress it up. What these reviews are for is the shape that emerges when several of them sit side by side: whether a stated 65% keeps landing near 65%, whether the gap between our number and the market's keeps costing us, whether a competition we read badly this week is still being read badly in three. Those are answerable, and they are answerable in public, which is the only reason to publish the bad figures alongside the good ones.
Every figure here comes from predictions published before the match day — never from a re-run after the results were known. Results come from public fixture feeds. Every period we have reviewed → How the model is measured →