Leeds vs Brentford: when our confidence score sits below our own probability
Published
There is a number on our Leeds United vs Brentford page that looks like a typo. In our 26 August model run the fixture is priced 38.4% Leeds, 27.1% draw, 34.5% Brentford — and the confidence score attached to it is 30.7%, lower than the probability of the outcome it is expressing confidence in. When a prediction confidence score sits below the model’s own top probability, something specific has happened, and it is worth explaining rather than hiding, because it is one of the few places where our model tells you it is arguing with itself.
What the confidence number actually measures
It is not a second probability. Our confidence score starts as the highest of the three outcome probabilities and is then damped by a documented heuristic: multiplied by 0.9 when the top pick is the draw, which our secondary signal cannot corroborate either way, and multiplied by 0.8 when that secondary signal disagrees with the pick.
Check the arithmetic here: 0.8 × 38.4 = 30.7. That is exactly the published figure. The damping has fired, and it has fired for the disagreement reason, not the draw reason.
So the honest reading of the Elland Road number is: our primary model makes Leeds narrow favourites, and our secondary model does not.
Why the two disagree about this particular fixture
FootInsights runs two things over the same match record. The primary one is a Dixon-Coles goals model, which fits an attack and a defence parameter for every club from the full history of scorelines, with older matches down-weighted. The second is a football-adjusted Elo ledger, which cares about who won and by how much — it applies a fixed multiplier to the rating swing that rises with the goal margin: 1.0 for a one-goal win, 1.5 for two, 1.75 for three.
Now look at what these two clubs actually did on the opening weekend.
Brentford beat Tottenham 3-0. Three goals through Keane Lewis-Potter, Vitaly Janelt and Michael Kayode, against a side that failed to register a first-half shot on target, per Sky Sports. In our Elo ledger that carries the 1.75 multiplier — the maximum band — so Brentford’s rating moved by nearly twice what a narrow win would have produced.
Leeds beat Nottingham Forest 1-0. One goal, an 88th-minute Anton Stach free-kick, in a game where Forest shaded the expected goals 0.65 to 0.49 according to Leeds’ own match report. Three points, and a 1.0 multiplier — the smallest rating movement the ledger offers for a win.
Identical results in the table. Very different results in an Elo book. That is the whole disagreement: the margin-weighted signal now leans Brentford, while the goals model, which is looking at far more than one weekend and weights Elland Road’s home effect, still leans Leeds by a fraction under four points.
What the disagreement is worth, and what it is not
It is worth treating the 38.4% as soft rather than firm. A four-point edge that only one of two signals supports is the definition of a fixture where the sensible position is “genuinely unsure”, and the confidence score exists to say so on the card rather than in a footnote.
It is not worth reading as a hidden lean towards Brentford. The damping does not move a single probability; the three outcomes still sum to 100 and Leeds are still the model’s pick. Confidence is a flag on the pick, not a competing forecast — and it is explicitly a first-pass heuristic, not a calibrated quantity. That limitation is stated on our how it works page and it will stay stated until there is enough settled data to fit a real one.
The part neither model can see
Both signals are built out of goals and results, and neither reads a team sheet. Squawka’s preview lists Mateo Joseph and Wilfried Gnonto out for Leeds, with Gabriel Gudmundsson and Ilia Gruev doubtful, while Brentford are without Antoni Milambo and Sepp van den Berg. Anton Stach and Ethan Ampadu, the two most direct contributors to the Forest win, look set to keep their places.
None of that is in the 38.4%. It is not weighted lightly — it is absent. If you think two attacking absences matter more at Elland Road than one defensive and one midfield absence do for the visitors, that is a view the model has not expressed and cannot contradict, and you would be adjusting a number that already carries an internal disagreement.
How to read the card on Sunday
Three things, in order. The published probabilities are the model’s answer and they will re-run before kick-off — the match page carries the current version, not the one frozen into this paragraph. The confidence figure tells you whether both of our signals point the same way; at 30.7% against a 38.4% pick, they do not. And how any of this has actually performed once matches settle belongs on the track record, which is the only place we publish scoring numbers.
Two teams on three points, separated in the table only by the margins of their opening wins, and a model that cannot make its own mind up. That is a more useful thing to know before kick-off than a confident-sounding pick would have been.