Flamengo vs Botafogo: the derby column our model does not have
Published
Round 25 of the Brasileirão sends Botafogo across Rio to the Maracanã on Sunday 30 August, kick-off 16:00 Brasília time, per Flamengo’s own match information. Our model’s 20 August run makes it 60.1% Flamengo, 22.4% draw, 17.6% Botafogo, and the live figure sits on the match page. The obvious question about a number like that in a fixture like this one is whether the model knows it is a derby. It does not. Do prediction models account for derbies at all, and does it matter here? Worth working through, because the honest answer is more interesting than either “yes” or “no”.
The number was computed before Flamengo lost
Two things have happened since that model run, and both belong at the top rather than buried.
Cruzeiro came from behind to beat Flamengo 2-1 at the Mineirão in round 24, as Yahoo Sports reported in its round-up of the round — Cruzeiro’s fourth straight league win, a run matched only by leaders Palmeiras, and Flamengo’s first Brasileirão defeat since 23 May. Botafogo, meanwhile, had not yet completed round 24 at the time of writing; their fixture against Athletico-PR was scheduled for Monday 24 August at 8pm Brasília, per the same report.
So the 60.1% on this page was produced by a run that had not seen either result. That is not a flaw to apologise for so much as a fact to state: prose freezes, and the embed above re-renders at every build. If the number on the match page has moved by the time you read this, the defeat is one of the reasons.
What the 60.1% is built from
Our team ratings, updated 23 August, put Flamengo 2nd of 20 in the Brasileirão on 1717 Elo, ranked 1st in the division for attack and 2nd for defence. Botafogo sit 5th on 1600, ranked 3rd for attack and 8th for defence. A 117-point rating gap, plus home advantage, plus a competitive attacking side on both ends is what produces a 60% favourite who is nonetheless expected to concede: 56.3% for over 2.5 goals and 53.9% for both teams to score.
None of those inputs contains the word clássico. The parameters are estimated from goals scored and conceded, match by match, with no marker for who the opponent is beyond their own rating.
Do prediction models account for derbies?
Some do, by adding a term that shrinks the favourite’s edge in designated fixtures. Ours does not, and the reason is a sample-size problem rather than a philosophical one.
A derby term has to be estimated from derbies. Flamengo and Botafogo meet roughly twice a season in the league. Even pooling every derby in the division, you are asking a model to learn a systematic effect from a category defined by hand, where the fixtures inside it differ enormously — a Rio clássico at the Maracanã and a mid-table local fixture are not obviously the same phenomenon, and treating them as one is how a parameter ends up describing whoever happened to win recently.
The things people actually mean when they say derbies are different — a full Maracanã, a squad rotated around a cup tie, a game that turns on a sending-off because the tempo was absurd from minute one — mostly show up as variance, not as a shift in who is likely to win. Our model has no channel for any of it, and would need years of clean evidence before it could justify one.
What would have to be true for 60.1% to be wrong
Rather than assert that the derby effect is small, put a size on it. Approximate the published split with an independent-Poisson model — crude back-of-envelope working, not the model’s internals — and it sits near 1.97 expected goals for Flamengo and 0.97 for Botafogo. Now suppose the occasion flattens the favourite:
- Flamengo’s scoring rate cut by 10%: 60.1% → 56.6%
- Cut by 20%: → 53.0%
- Cut by 30%: → 49.8%
- Or instead, Botafogo’s rate raised by 20%: → 56.0%
- Raised by 30%: → 53.8%
To turn this fixture into a coin flip you need to knock roughly a quarter off the best attack in the division, or add a third to the visitors’, purely because of what the fixture is called. Both are large claims. A derby effect worth three to six points off the favourite is defensible and would leave Flamengo comfortably ahead; one worth twenty is not something anyone has evidence for.
That is the useful conclusion. The model’s blind spot here is real, it points in a known direction, and it is small enough that it does not change the reading of the game — which is exactly the kind of limitation worth naming out loud rather than letting a reader discover it.
How we will grade it
Whatever the embed shows on the day is what gets scored, with its model version attached, in the public track record — including if a Maracanã derby does something no rating file anticipated. That is the whole point of publishing the number in advance instead of explaining it afterwards. Botafogo’s team page carries their ratings as the season moves them, and the rest of round 25 is on the Brasileirão hub.