How the Golmetria Index works

Our ratings, probabilities, and value layer — what the model does, what's live today, and how we keep ourselves honest.

The Golmetria Index is a rating-and-forecasting system for Brazilian club football. Its job is prediction, not merit: it estimates what will happen, not who deserves to rank highly. Every output is a probability, and probabilities are sometimes wrong — that is the nature of the thing, and the reason a betting market exists at all.

It is independently designed. It draws on public, well-established ideas — Elo ratings, the expected-goals literature, negative-binomial scoring models — and implements them from scratch on our own data. It does not use, copy, or republish any proprietary model's ratings.

This page is the canonical, versioned description of the model. It is analysis, for adults 18+, on licensed (bet.br) operators — never betting advice, never a guarantee.

The data

The model reads finished match results and expected goals (xG) from a licensed data provider, and the fixture calendar live — kickoffs are never hard-coded. xG is the key input: it strips out finishing luck and goalkeeping variance, so it describes how a match should have gone better than the scoreline alone. Where xG is unavailable (lower divisions, some state matches), the model falls back to goals and widens its uncertainty.

The Índice — strength rating

At the core is the Índice, an Elo-style rating where 1500 is the average club, updates are zero-sum between the two teams, and every number is contemporaneous — computed only from information available at the time, with no hindsight.

After each match a club's Índice moves toward the result, scaled by how surprising it was. Four choices shape that update:

• An xG-and-goals blend. Performance is measured mostly by xG, partly by goals (we weight xG around 0.6). xG is more predictive; goals still carry information, so we keep both.
• A harmonic margin. The second goal counts for less than the first, the third less than the second. A 4–1 is not twice as meaningful as a 2–1 — teams ease off once a game is decided.
• Match-importance weights. A league match carries full weight; friendlies and rotation-heavy state group games carry less; cup and continental knockouts carry more.
• Home advantage, applied as a rating bonus to the home side and removed at neutral venues.

Newly promoted clubs, which we know little about, update faster for their first stretch of top-flight matches and then settle to the normal pace.

Pendor — the over/under lean

The Índice answers who wins; it says nothing about how many goals. On its own the model would expect the same total in every match and only shift it between the sides by strength. Pendor is the second rating that fixes this — a slow-moving lean tracking whether a club's games tend to run higher- or lower-scoring than strength alone implies, which feeds the over/under. It moves gently: goals are noisy, so it regresses hard toward neutral and only a persistent tendency shows through. Calibrated on fifteen seasons of history, turning it on measurably sharpens the over/under.

From ratings to probabilities

Two ratings are not a forecast until they become a distribution over scorelines. For each fixture the model projects total goals (a league scoring baseline tilted by Pendor), splits them between the sides by strength, and builds a score matrix — the probability of every exact result — using a negative binomial distribution rather than Poisson. Negative binomial is over-dispersed: it allows the extra draws and the occasional blow-outs that real football produces and Poisson does not. On top of that we apply a small Dixon-Coles correction so the two scores aren't treated as strictly independent — independence under-counts draws. Calibrated on this season it is a gentle nudge (the negative-binomial base already lifts the draw rate close to where it should be), but it is the textbook fix and it is in the live model. Summing that matrix gives the published markets: 1X2, over/under, both-teams-to-score.

Season projection

To turn single-match forecasts into a title race, Libertadores chase, or relegation fight, the model runs a Monte-Carlo simulation of the remaining fixtures thousands of times. It runs hot — within each simulated season the ratings update as results occur, so streaks compound realistically — and reports each club's probability of every final position. That is the finish-position heatmap and the projected table on the Index page.

The value layer

This is the part most models keep private, and the reason Golmetria exists. For a given market we compare our probability with the market's fair probability:

1. Take the bookmakers' odds and remove the margin (de-vig) to recover the market's fair line.
2. Edge is our probability minus that fair probability.
3. Expected value is computed at the best price actually available — the price you could take, not an average — because value only exists at a price you can get.

When an edge clears a threshold we flag it, and we publish the model probability, the price, and the edge together — so you see the reasoning, never a naked "tip." Any mention of stake discipline is framed as fractional-Kelly analysis, never a directive and never an amount.

How we keep ourselves honest

Transparency is the brand, so the model is built to be graded in public. We track, and will publish as the record accumulates:

• Closing Line Value (CLV) — did a flagged play beat the closing price? It is known at kickoff, before any result, and is the single best signal that an edge is real rather than luck.
• Brier score and log loss — how accurate our probabilities are.
• Calibration — when we say 60%, does it happen about 60% of the time?

Honest reporting of losing stretches is the point, not an embarrassment. The methodology is versioned and dated, and changes are logged.

What's live today, and what's next

We publish only what is built. Live now: the Índice (Elo with the xG-and-goals blend); the Pendor over/under lean; the Dixon-Coles negative-binomial score matrix; the Monte-Carlo season projection and finish-position heatmap; the value layer (de-vig → edge → expected value at best price); and the CLV/accuracy tracking that will populate once the season resumes and bookmakers price fixtures.

On the roadmap: a squad-market-value prior with transfer-window reversion (which would add an attack-vs-defence split to Pendor); richer home advantage (travel and altitude); Copa do Brasil and continental competitions; and the published track record. We would rather ship a smaller honest model than claim a bigger one.

Limitations and responsible use

It is a model, not a crystal ball. A positive expected value still loses individual bets — the value layer manages variance, it does not remove it. Cross-league comparisons rest on few matches and carry wider error bars, and squad changes, injuries, congestion, and motivation move single games in ways no model fully captures.

Everything here is statistical analysis, not betting advice, intended for adults 18+, on licensed (bet.br) operators only. Betting is not a source of income. Bet responsibly — and if it stops being fun, step away and seek help.