xgEdge

Methodology

How the model is compared with the market

Every row in the Betting section sets a probability from our model beside one bookmaker's price for the same outcome. This page defines the words the section uses, explains each number on that row and the arithmetic between them, then states the one rule that decides which rows are signals, and at what moment. A signal is a disagreement the rule lets through, not a recommendation. The football model itself is explained once, on How the model works.

Glossary

One word, one population

The section uses these words, and no others, for the sets of rows they name. Each stands for one population, and none is used for a population it does not name.

Evaluated price
Any row the model has priced against a bookmaker: one selection, one price, one probability.
Model probability
The model's estimate of an outcome's chance, read off its projected scorelines. Not used for: the market's view, a recommendation.
Market probability
The bookmaker's probability with its margin removed, scaled over a complete book in which every outcome is priced. Where one is not, the raw 1 / odds stands in and is marked as including the margin.
Model price
1 / model probability: the price at which a stake breaks even if the model's probability is right. The model does not know the right price; this is what its probability implies.
Model-implied edge
Model probability × market odds − 1: what a stake of one would return on average at the offered price if the model is right. It is not observed return. Not used for: the disagreement, a result anyone has seen.
Disagreement
Model probability minus market probability, in percentage points. It says the two disagree, not which is right. Not used for: the model-implied edge.
Positive edge
An evaluated price whose edge is above zero, nothing else implied. Not used for: anything qualified.
Signal
An evaluated price that satisfies the canonical rule at a stated moment; its status word is Signal. It is a disagreement the rule lets through, not a recommendation and not evidence. Its model-implied edge, model probability × odds − 1, is the model's claim, not an observed return. Not used for: a recommendation, evidence, “value”, “opportunity”, “pick”.
Recorded signal
A signal in the retrospective record, settled or pending, at its own evaluation moment. On the Performance page, a selection always means a recorded signal.
Evaluation moment
The time of a row's last model evaluation before kickoff, when its probability and price were stored. The record asks the rule about each selection at that moment, never at today's clock. Not used for: kickoff, the present.
Settled / Pending
A recorded signal whose match has a result is settled: won, lost or void. One whose match has kicked off and has no result yet is pending, which is neither settled nor a loss. Not used for: a loss.
Stake / Units
One unit on every recorded signal, equal and never varied. Units are the stake's own measure: a result of −3.5u is three and a half stakes lost.
Fresh / Stale price
The price is within, or beyond, its tier's allowance at the assessment moment. Not used for: “current”, “live”.
Full data / Limited data
The match carries no model warning, or carries one. Not used for: confidence.
Forecast skill: Beats base rate / Doesn't beat base rate
The backtest's skill figure for the market is above zero, or is not: the model predicts the market better than a historical base-rate forecast, or does not. This does not mean it beats bookmaker prices. Not used for: “validated”, beating the bookmaker, a measure of profit.
Evidence
Historical, empirical support for signals like this one, and nothing else: on a card, the forward status of its category — market selection × model-implied edge band — in the live signal ledger, then the retrospective record of past signals in the same category, labelled retrospective. A category stays Not yet validated until its live signals reach the promotion floor. A model probability, a disagreement or a model-implied edge is a model output, never evidence. Not used for: a model output, a status word, a recommendation.
Brier score
The squared distance between each forecast probability and what happened, averaged over the same matches, from 0 for a certain and correct forecast upward: lower is better. Not used for: a measure of profit.
Calibration gap
How often the signals happened less the model's average probability for them, in percentage points: below zero the model was overconfident, above it underconfident.
CLV
How far the market moved toward a signal between the moment it was recorded and the same bookmaker's price about an hour before kickoff: above zero it moved toward the selection, which is a weaker sign than a sharp market's move and is not a profit. Not used for: a result, a profit.
Retrospective evidence
The record of what today's rule finds in past prices, each read at its own evaluation moment; nothing in it was published at the time, so it cannot demonstrate an edge. Not used for: a live record.
Prospective validation
The ledger of signals as they were published live, settled once their matches are played; only it can show an edge, and only once it reaches the promotion floor.
Goal calibration
Whether the goals the model predicts match the goals scored, read two ways: the slope of the real goal difference on the predicted one, and goals scored over goals predicted. Both are 1.0 for an honest model.

Value, opportunity, pick and best are retired from the section's pages. Each had come to mean two or three of the populations above at once; a signal is what they were reaching for, and it is the only word now used for it. Fair odds, model edge, qualified selection and probability gap are retired with them: the words are Model price, Signal and Disagreement. “Proven skill” is retired as well: the backtest measures skill against a base rate, not against the bookmaker, and the words for it are Beats base rate and Doesn't beat base rate.

The numbers

Model probability, model price, implied probability and overround

The model probability is our model's estimate of how likely an outcome is, written to the price row the last time the model was run. No page in this section works a probability out any other way.

The model price is the price the model breaks even at if its probability is right: one divided by the model probability. A 50% outcome has a model price of 2.00. The model does not know the right price; this is only what its probability implies. If the market offers more than the model price, the model's probability says the price is too long; if it offers less, too short.

The implied probability is the same conversion run the other way on the bookmaker's price: one divided by the odds. Odds of 2.50 imply 40%.

The overround is how much more than 100% a bookmaker's prices add up to across every outcome of a market. It is the bookmaker's margin: a three-way market whose implied probabilities sum to 104% carries a 4% overround.

The margin

How the margin is removed, and why the ranking does not use it

To compare the model with what the market itself thinks, the margin has to come out. We use the multiplicative method: each outcome's implied probability is divided by the market's total, so the outcomes sum to exactly 100% again and each keeps its share.

It only runs when every outcome of the market is priced. With one leg missing the total would be short and the margin understated, so in that case the market figure is shown as it is and labelled as including the margin.

The margin-removed figure is only used in the model-vs-market bars, the disagreement beside them and the Model vs market calibration on the Performance page, which scores the market's probabilities with the margin removed beside the model's. The model-implied edge, and everything ranked, filtered or let through by it, uses the bookmaker's actual price with the margin still in it, because that is the price on offer.

Edge and the gap

Model-implied edge and disagreement, side by side

A row carries two numbers that say how far the model is from the market, and they are not the same number. The model-implied edge is the model probability multiplied by the market odds, minus one: what a stake of one would return on average over many bets like this one, if the model's probability were exactly right. It is the model's claim, not an observed return. A 55% outcome at 2.00 has a model-implied edge of +10%; the same outcome at 1.70 has one of −6.5%.

The disagreement is the model probability minus the market's probability, in percentage points. The market's probability is the margin-removed figure when every outcome of the market is priced, and the implied probability, with the margin still in it, when one is not; the row says which. A 55% outcome against a market at 50% has a disagreement of +5.0 points whatever the price. It says the two disagree, not which is right.

The edge is measured against the price on offer, so it ranks, filters and qualifies; the gap is measured against what the market thinks, so it explains. Two prices with the same gap can carry different edges, because the margin in them differs, and the price is what is paid.

A positive edge is a statement about an average, not about one match. A +10% edge on a 55% outcome still loses 45 times in 100. Over a small number of bets the results are dominated by that variance, and a run of losses is not evidence the edge was wrong, just as a run of wins is not evidence it was right. It also rests entirely on the model probability, which is an estimate with error in it.

Price needed

The price an edge needs

Since the model-implied edge is probability × odds − 1, the price at which it reaches 5% is one plus the threshold, divided by the model probability: (1 + 5%) ÷ probability. It is the model price scaled up by the threshold, and it needs no market price at all.

It says how far a price is from the line rather than only whether it is over, which is the figure to read when an edge is close. The worked example below ends on it.

Worked example

One row, worked through

A made-up match priced by the model as this page renders, against a complete three-way book (Home win 2.15 · Draw 3.40 · Away win 3.60). We follow Home win through every figure an Overview row shows.

The match's ratings are constructed rather than fitted, so the example reads the same under every model version; real matches are priced by version 1, whose settings are under The model's side below.

  1. Model probability 49.7%
  2. Model price 1 ÷ 49.7% = 2.01
  3. Implied probability 1 ÷ 2.15 = 46.5%
  4. Overround 1 ÷ 2.15 + 1 ÷ 3.40 + 1 ÷ 3.60 = 103.7%, so 3.7%
  5. Market, margin removed 46.5% ÷ 103.7% = 44.9%
  6. Disagreement 49.7% − 44.9% = +4.8 pp
  7. Model-implied edge 49.7% × 2.15 − 1 = +6.7%
  8. Price needed for a 5% model-implied edge (1 + 5%) ÷ 49.7% = 2.11

So Home win would show a model-implied edge of +6.7% on the Overview and a disagreement of +4.8 pp against the market's margin-removed figure, and its model-implied edge reaches 5% at a price of 2.11. The model-implied edge is measured against the price on offer; the disagreement is measured against the market with its margin taken out.

Each result is computed from the unrounded figures and then rounded the way the Overview rounds: probabilities to one decimal, odds to two, edges signed to one decimal. Redoing a step from the rounded numbers shown can land one digit off.

Two answers to one question

The same match, two more of its prices: Home team to score at 1.30 and Home team 2+ goals at 2.25. Both answer how many goals the home side scores, and both pass every other criterion of the rule, so the rule keeps one.

  1. Home team to score · 1.30 82.7% +7.5% Outranked
  2. Home team 2+ goals · 2.25 49.7% +11.9% Kept

The rule keeps Home team 2+ goals (+11.9%) and says Outranked on Home team to score (+7.5%). The kept answer is the larger expected value, not the likelier outcome: Home team to score is the more probable of the two and is still not the match's answer to that question. It stays in All prices with its figures and is not a signal.

The rule

What a signal is

One rule decides, for one evaluated price at one moment, whether it is a signal. Every count, card, colour and record in the section reads that rule, and nothing tests a criterion of its own. The moment is stated: the live pages ask about now, the record asks about each selection's last evaluation before kickoff, and the rule is the same whichever it is. A signal is not a recommendation.

At the moment asked about, the price is a signal when all of these hold, in this order: the match has not kicked off; the price was captured no later than the evaluation that priced it, so the stored probability was really set against it; the price is fresh by the odds schedule's tier for that distance from kickoff; the match carries no model warning and both sides have a sample behind their rating; the model beats its base rate on the market; the model-implied edge, from the stored probability and price, is at least 5% and below 40%; and, among the match's prices that pass all of these, it is the answer to its question with the largest expected value — one answer per question per match.

Every market answers one of five questions, and these are the questions with the markets that answer each: who wins (Full-time result (1X2)); how many goals the home side scores (Home team to score, Home team 2+ goals); how many goals the away side scores (Away team to score, Away team 2+ goals); how many goals there are (Over 1.5 goals, Over 2.5 goals); whether both sides score (Both teams to score). Outranking is decided among prices that pass every other criterion: a price that fails any other criterion neither outranks another nor is outranked by one. Where two answers to one question tie on expected value at the stored precision, the one stored first is kept. An outranked price stays in All prices with its figures, says Outranked as a fourth status word, and is never a signal (Outranked, below).

Each criterion that fails is a reason, and a row says its three status words — Fresh or Stale price, Full data or Limited data, Beats base rate or Doesn't beat base rate — whatever else it fails on, and a fourth, Outranked, when another answer to its question is kept. Comparisons are made at the stored columns' 6 decimals, so a figure stored exactly on a line sits on it. No verdict is stored: it is computed when asked for, because one written down would be wrong the moment the clock moved.

Qualification is per answer: one per question per match. Where a match has qualified answers to more than one question, the cards show the one with the largest edge, then the earliest kickoff, while the board and the counts show them all. One card per match is presentation, and it changes no verdict.

Freshness

Fresh and Stale price: the odds schedule

Prices are fetched on a schedule tiered by time to kickoff, because the provider serves odds one match at a time and bills per request. Within 24 hours of kickoff a match's prices are refreshed every 3 hours; from there out to 72 hours, every 24 hours. Nothing is fetched further out: most bookmakers have not posted every market yet, and the requests would find nothing.

The model is re-run against the stored prices after every sweep, so a row's edge pairs each price with a probability evaluated against it.

A price is Fresh while it is within its tier's allowance and a Stale price once a refresh should already have replaced it: after 6 hours within the last 24 hours before kickoff, and after 48 hours before that. The tier is the one the match is in at the moment asked about, so a price a match two days out may keep for 48 hours is Stale at 6 on the day. Stale prices stay in All prices, with the word and the price's age beside the model-implied edge, and are never a signal.

The odds schedule's tiers
Time to kickoff Refreshed every Stale after
Within 24 hours 3 hours 6 hours
24 to 72 hours 24 hours 48 hours

Limited data

Full data and Limited data: the sample behind a rating

A club's rating is fitted from the shot data of its matches. When a side has fewer than 10 such matches behind it — a newly promoted club, a competition early in its season — the match carries a model warning naming the side and its count; when a side has no sample at all, the model prices the match on the competition's average for that side and the warning says so. In either case every row on the match says Limited data, and is never a signal.

Full data means neither: no warning, and a sample on both sides. It is a statement about how much the probability rests on, not about how sure the model is. The model has no uncertainty estimate of its own; it does not know about injuries, rotation, weather or anything else outside the shot data, and nothing on these pages should be read as a statement that one outcome will happen.

Skill

Beats base rate and Doesn't beat base rate: what the backtest found

The model was replayed over a season of real matches in 7 competitions, measured on 17 August 2026. For every market it priced, its probabilities were scored against what happened and set against the base rate — simply knowing how often that outcome happens — and the difference is the market's skill figure.

Beats base rate means the figure is above zero: the model beat the base rate on that market, by however little. It does not mean it beats bookmaker prices. It did on Away win, Home win, Home team to score, Home team 2+ goals, Away team 2+ goals, Away team to score. It did not on Over 1.5 goals, Both teams to score, Over 2.5 goals, Draw, and zero counts as not, since a model worth exactly as much as the base rate is worth nothing on it.

A market where the model doesn't beat the base rate stays in All prices so the comparison is complete, but every row on it says Doesn't beat base rate and is never a signal, whatever its model-implied edge. Doesn't beat base rate does not mean the market was never measured: it was, and the model did not beat the base rate on it. Beats base rate is not “validated”, not beating the bookmaker and not a measure of profit: it says the backtest's figure was positive on those competitions over that season, and no more. The Market overview carries the same split per selection.

Outranked

Outranked: one answer per question

Several of the markets on a match answer the same question — whether the home side scores and whether it scores twice both ask how many goals the home side scores — and a match gets one answer to each. Among the match's prices that pass every other criterion, the one with the largest expected value is kept; every other answer to the same question says Outranked, keeps its figures in All prices, and is not a signal. The kept answer is on the board, and an outranked row's details name it.

Outliers

The outlier line

A model-implied edge of 40% or more is treated as a likely data error rather than a price. A bookmaker's price that far from the model's is far more often a wrong price, a market suspended and not withdrawn, a line keyed against the wrong selection or a probability pinned by thin data than a bet. Such rows stay in All prices, with the model-implied edge printed as it is, and are never a signal.

The line is a cutoff on the quoted figure, not a judgment on any one row, and a row just under it is treated like any other: a signal's model-implied edge runs from 5% up to but not including 40%. Reading a three-figure model-implied edge in All prices as a real one is the mistake the line exists to prevent.

The odds

Where the prices come from

Prices come from one odds provider and are one bookmaker's, not the highest available and not a consensus of several. For every market we take the first of Bet365 → Paddy Power → Pinnacle that prices it, in that fixed order, and the bookmaker is named on the row.

The order is fixed rather than chosen per match, so no price is taken because it happens to disagree with the model. Bet365 comes first because it covers most of the competitions here; the others fill the leagues where it does not.

The record

How the record is kept

The signal record on the Overview, the Performance page and the homepage is one tally, kept by these rules. Each term below is the rule's own definition: the Performance page prints the figures, and this is what they mean.

What counts
Every evaluated price that was a signal at its evaluation moment, in every competition the site prices: a fresh price by the odds schedule's tier, full model data, a market where the model beats its base rate and a model-implied edge from 5% up to but not including 40%. It is the same rule the live board reads, asked at a different moment.
The moment
Each row is assessed at its own evaluation moment, with the probability and the price stored then. Nothing about a past selection is decided from today's clock, and a selection that qualified then counts whatever the price did after.
The price
The bookmaker's price paired with that evaluation, from the bookmaker named on the row. Nothing is re-priced with a later or a better price, and it is not a closing price (see The record's price). A row whose price was captured after the evaluation that was meant to price it is excluded and counted.
No backfill
A row written after its match kicked off is never counted, whatever it says. The record is of what the board would have highlighted before the match, and a row that could not have been on the board then cannot be in the record now.
Staking
An equal stake of one unit on every recorded signal, so a match with two signals carries two units. A win returns the price less the stake, a loss costs the stake, and a void returns it and still counts as staked.
Return
Units are the sum of those returns, and ROI is units divided by units staked, quoted only from 30 settled selections, overall and in every breakdown row. Under that it is mostly the last few results. Returns are gross: before any commission, tax or fee.
Regrading
A selection is graded from the score stored when its match is read, and graded again if a later reading of the match corrects that score. A regraded selection keeps its settlement time and carries the time of the regrade and the score it had been graded from. So far 236 of the board's settled prices have been regraded.
What is left out
Pending selections, which are neither settled nor losses; rows without a price; the excluded pairs above, which are counted; and the daily combinations, a different bet on the same matches, reported on their own pages.

The Performance page can narrow the record by period, market family, competition, edge band or odds, and a slice is descriptive only: it reports what happened on that subset, and the 5% threshold is set by the rule, never tuned to what a slice returned.

The 95% interval resamples matches rather than selections, 2,000 times from seed 20261102, so selections on one match move together.

11 voids are staked and counted. 929 settled rows are left out of every figure: their price was captured after the evaluation it was paired with.

Retrospective: today's rule asked about each selection at its last evaluation before kickoff, not a list published at the time. A quoted historical price may not have been available to you. Computed .

Settlement

Pending, won, lost and void

A recorded signal is pending from kickoff until its match has a result. Results are graded in one sweep once they land, so a match postponed by a week keeps its selections pending for that week. A pending selection is neither settled nor a loss: it adds nothing to the record until it is graded.

A selection is won when the outcome it names happened, by the same reading of the scoreline that priced it, and lost when it did not. A selection on a match that was never played — postponed without a new date, cancelled or abandoned, and recorded as void — is void: it was a price on a match that did not happen, so it carries no verdict about the price. A void returns the stake and still counts as a settled, staked selection.

A verdict, once written, moves for one reason only: a later reading of the match corrects the score it was graded from, and the selection is graded again from the corrected score with its settlement time kept and the regrade recorded on it (see How the record is kept). Nothing else re-grades a row, so the record is an archive and not a running estimate.

The record's price

What the record's price is, and is not

Every selection in the record is settled at the price stored with its last evaluation before kickoff: the bookmaker's price the model's probability was set against, captured at most its tier's allowance earlier, which within the last 24 hours is under 6 hours. It is the price a reader of the board would have seen beside the edge at the time.

It is not a closing price. There is no capture at kickoff, and nothing is re-priced with a later or a higher price. A price may have moved, been limited or been withdrawn between the capture and kickoff, and the record cannot say which way, so the Performance page says this beside its figures rather than implying a closing line.

Correlation

Why two selections on one match are not two experiments

A match can carry more than one signal — a home win and the home side to score, say — and each is a selection staked at one unit. They are settled by the same ninety minutes, so they tend to win and lose together, and a count of selections overstates how many independent results the record holds.

That is why every selection count in the record carries the distinct match count beside it, and why the smaller figure is the one to read a return against. The equal stake is kept rather than split across a match's selections, because splitting it would let the record's size depend on how many markets a bookmaker happened to price.

Two calibrations

Model vs market, and Market-specific calibration: what each is for

The Performance page opens with two calibration sections that answer different questions over different populations, and the words on the page keep them apart. Model vs market: 1X2 is the benchmark: the model's home, draw and away probabilities beside one bookmaker's, with the margin removed, on exactly the same matches, each scored against what happened. It asks whether the model's probabilities meant more than the market's, which is the fairest test a reader can put to a model whose probabilities are then set beside prices.

Market-specific calibration is the model alone, across every betting market the site supports: each selection the model priced on a played match, from the last projection before kickoff, whether or not a bookmaker quoted it. It asks whether a stated probability meant what it said, market by market, and it has no market side to compare with, because the bookmaker does not price every selection the model does.

The two are computed by different services over different samples and are not to be read against each other. A Brier score in the first is taken per match over the three outcomes together and runs from 0 to 2; in the second it is taken per selection on its own and runs from 0 to 1, so the figures are on different scales even where the words are the same. The calibration of the signal record, further down the page, is a third reading, over the record's rows alone, and is defined in the section that follows these.

The model's side

How the model's probabilities are produced and frozen before kickoff

Each probability is the football model's, explained once on its own page: expected goals for and against from each club's recent shots, fitted into an attacking and a defensive rating per club, turned into two expected scoring rates for the match and run through a grid of every scoreline. The home, draw and away probabilities are the three regions of that grid summed, so they come to exactly 100% before any price has been seen.

Each time the odds sweep prices a match, the model's probability current at that moment is stored on the evaluation row beside the bookmaker's price, with the time of the evaluation. The comparison reads the last such row written before kickoff, and nothing rewrites a played match's row: the probability and the price are frozen together on one record, so neither side can be re-priced with hindsight, and what is scored is what the site published, not a model re-run later.

That also means the comparison measures the model as it stood on each day rather than one fixed version of it. The model's inputs last changed on 26 August 2026, when version 1 took effect; the Performance page counts the matches priced before that change, and every change is dated in the history below. A match the model priced with a warning, or with a side it had no sample for, is included and counted rather than removed: the question is what the model said, warned or not.

The model in force is version 1, since 26 August 2026. Each club's ratings are fitted on its latest 20 matches with expected goals, on a half-life of 107 days, with shot volume weighted 0.5 beside expected goals; a club with few matches is pulled toward average by 5 matches' worth of an average side; and each competition's baseline is its expected goals, with no correction toward the goals it scores.

The market's side

How the market's probabilities are calculated and the margin removed

A decimal price is turned into an implied probability by taking its reciprocal: a price of 2.50 implies 40%. Over the three outcomes of a match the implied probabilities sum to more than 100%, and the excess is the bookmaker's margin, the overround.

The margin is removed over all three outcomes at once, by the multiplicative method the section uses everywhere (The margin, above): each implied probability is divided by the three's total, so they sum to exactly 100% again and each keeps its share of the book. Prices of 2.50, 3.40 and 2.90 imply 40.0%, 29.4% and 34.5%, a total of 103.9%; with the margin removed they read 38.5%, 28.3% and 33.2%. The arithmetic is one function shared with every row on the Overview and the match page, so the calibration's market probability and the board's margin-removed figure cannot differ.

Only a full book is used: the three prices from one bookmaker, every one captured before kickoff. A match with a 1X2 price missing, priced after kickoff or split across bookmakers is left out of the sample entirely rather than de-vigged on a short book, which would understate the margin. The bookmaker is the first in the fixed order of preference (Bet365 → Paddy Power → Pinnacle) with a full book, so a match is one book and one price and never a consensus; the Performance page prints how many matches each bookmaker supplied. Methods that assign more of the margin to the longer-priced outcomes are not used.

The price

Which price is used, how far before kickoff it was captured, and why it is not a closing price

The market's price in the comparison is the last pre-kickoff market price: the price stored on the model's last evaluation of the match before kickoff. It is captured by the odds sweep, which refreshes a match's prices on a schedule set by its distance from kickoff (Freshness, above): within 24 hours of kickoff a price is refreshed every 3 hours, so the last capture typically sits some hours before the match rather than minutes.

The Performance page prints how far before kickoff the price was captured — the median, the quartiles and the whole distribution in bands — rather than one figure. The spread is wide, and some matches were last priced more than a day out; they are in the sample as they are, and the reader can see how many.

It is not a closing price, and the page never calls it one. The odds provider publishes no closing line, and a price captured hours before kickoff has not seen the late money, the team news or the market's final correction. The market side of the comparison is therefore somewhat weaker than the price at kickoff would be, which, if anything, favours the model. A price captured closer to kickoff would be its own benchmark over its own paired sample, published under its own name; replacing the one the page reads would be a dated entry in the history below, never a silent change.

Scoring

Bands, floors, Brier score, log loss and the average calibration gap

Each forecast is a probability and an outcome that did or did not happen. The reliability charts group the forecasts into 10 equal-width bands of the probability scale, closed on the left so that a probability of exactly 100% falls in the last, and plot each band's mean probability against how often those outcomes happened. A point on the diagonal is calibrated; above it the forecasts were too cautious, below it too confident. The pooled view takes all three forecasts of every match; the home, draw and away views take one each, so a pooled chart cannot hide that one outcome is priced differently from the others.

Two floors. A band is drawn from 30 forecasts, counted per source, so it may be drawn for the market and not for the model, and the table shows both counts either way. The section as a whole is read from 300 paired matches; below that it prints the count so far and nothing else.

The Brier score is taken per match over the three outcomes together: the sum of the squared differences between each probability and the outcome's one or nought, from 0 for a certain and correct forecast to 2 for a certain and wrong one, averaged over the matches. The log loss is minus the natural logarithm of the probability given to the outcome that happened, clamped just above zero so that one impossible forecast costs a great deal rather than everything, averaged the same way. Lower is better for both, and both are computed for the model, the market and a base rate on exactly the same matches in one pass. The base rate gives every match the sample's own home, draw and away frequencies, measured on the sample itself, so it is the strongest version of the reference and could only have been known with hindsight.

The average calibration gap under each chart is the expected calibration error, ECE: each band's distance between its mean probability and its observed frequency, weighted by the band's share of the forecasts and summed over all 10 bands, the undrawn ones included. It depends on the choice of bands and summarises the diagram; it is not a score of accuracy, and two forecasters with the same gap can differ widely on the Brier score.

The sample

Which matches are compared, and how the sample grows

The sample is every played match with a final score whose three 1X2 rows carry a price fetched before kickoff from one bookmaker and a model probability evaluated before kickoff. Nothing is chosen by competition, by the size of an edge or by whether the model and the market agreed; a void match, an unscored one and any match with a short book are out, and that is the whole of the selection. One sample serves the model, the market and the base rate, so nothing is scored on matches another side lacks.

The sample runs from the first match whose evaluation stored the price beside the model's probability to the last played match, and the Performance page prints the period. It grows as matches are played, so every figure on the page moves, and the page states when it was last computed: a figure quoted from it is a figure as of that date.

Three forecasts on one match are not three matches. The pooled view's forecast count is three times the match count, and the match count is the one to read the sample's size by. No confidence interval is printed, because the three forecasts of a match are dependent and the site has no method yet that respects that; a small difference between two scores is not a finding.

Qualified calibration

What the calibration of the signal record measures

The Performance page scores the model twice, over two populations. Model calibration reads every selection the model priced on a played match, from the last projection before kickoff, whether or not any price was quoted on it. The calibration of the signal record reads the record's own rows only: the settled selections that were signals at their last pre-match evaluation, each with the model probability stored at that evaluation and its stored result. A void is counted and not scored, since a match that did not happen gives no outcome to be right about; a pending selection never enters.

Over the decided selections — won or lost — the section prints the model's average probability, how often the selections happened, the difference between the two in percentage points, and the Brier score, the mean squared distance between each probability and what happened. The base rate beside the Brier score is the same reference the overall calibration uses, computed on exactly this subset: a forecast that gave each selection its own frequency over these rows, which it could only have known with hindsight.

The two selection counts are the record's, so the section and the record can never disagree about how many rows they read. Selections from the same match are settled by the same ninety minutes and may be correlated, so the selection count overstates the independent results behind the reading. A signal is one the model priced above the market, so the subset is chosen by the model's disagreement with the bookmaker; a reading over it describes the recorded sample and is not a fixed property of the model.

CLV

How the market's move after a qualified price is measured, and against which price

Every price the qualification rule qualifies is entered in a ledger the first time it qualifies — the bookmaker, the price, the model's probability and the market's probability with the margin removed, as they stood at that evaluation — and the entry is never rewritten. Once the match has kicked off, each entry is set beside the same bookmaker's price on the same selection near kickoff, and CLV measures how far the market moved between the two.

The primary form compares probabilities with the margin removed: CLV = p_close ÷ p_entry − 1, where p_entry is the selection's margin-free probability when it qualified and p_close the same at the close, each from that one bookmaker's whole book. Above zero, the market moved toward the selection after it qualified; below zero, away from it. The secondary form compares the raw prices: price CLV = odds at entry ÷ odds at close − 1. It is the familiar one, but it carries any change in the bookmaker's margin along with the market's move.

The close is Bet365, about an hour before kickoff, not a sharp closing line. More exactly, it is the entry's own bookmaker — Bet365 for most matches — captured by a run in the 60 minutes before kickoff, and counted only when captured between 0 and 90 minutes before it. The odds provider publishes no closing line, and one bookmaker's price an hour out has not seen the last of the late money or the team news, so a positive CLV against it is a weaker finding than the same figure against a sharp book at kickoff.

Nothing is filled in from elsewhere. An entry whose selection was not captured, was captured outside the window or was captured from another bookmaker gets no CLV, and is counted by its reason. One whose book was short at either end — an answer to the question unpriced — has no margin-free probability to compare: it carries the price form only, and is counted too.

The signal ledger, public rule: CSV · JSON Every signal as it was written, with its close and its result. Shadow rules are not included.

Model vs market, 1X2: CSV · JSON Every band of the comparison, with its counts and the method's metadata.

History

Method history

Every material change to the method — the football model, what “qualified” means and how the model is compared with the market — dated, newest first. A change here changes the probabilities the model publishes, which selections the record counts or what a comparison measures, so the figures published before and after it are not comparable. An entry records a change; it is not a claim that any figure got better or worse.

  1. Effective

    Publication: research note, first edition

    Before: the forensic report on the retrospective record lived in the project's working files, and no page set out what is failing, why, and what is being tested next.

    Now: the Research note publishes it, dated, with its database, as-of moment and rule version at the top: what is failing, why it probably is, the validated tests apart from the exploratory cuts, the data-quality census, what remains unknown, the hypotheses and the pre-registered shadow rules with their promotion test. Each later edition is published beside it at its own address, never over it.

    Why: a diagnosis kept private cannot be checked, and a frozen reading has to say it is frozen so it is never read as live.

    Recalculated: nothing. The edition's figures are the research script's on a copy of the production database as of 29 October 2026; the live figures are on Performance.

  2. Effective

    Publication: Model vs market names its leader

    Before: Model vs market: 1X2 printed both scores and their difference and named no leader.

    Now: it leads with Market leads, Model leads or No clear leader yet, read from the paired per-match Brier and log-loss differences: a side leads only when both differences favour it and both match-resampled intervals exclude nought.

    Why: two scores side by side invite a reader to name a winner from a difference that may be noise. The leader rule names one only when the data does.

    Recalculated: nothing. The scoring is unchanged; the comparison's figures are the same, with their intervals added.

  3. Effective

    Publication: evidence on every signal

    Before: a signal's card carried the model's figures and the market's, and nothing on how signals like it had done.

    Now: every card carries an Evidence row: its category's forward status — Not yet validated until the category reaches the promotion floor in the forward ledger — and the retrospective cut of similar past signals, the same market selection in the same model-implied edge band, with its count, its matches, its ROI and interval and its calibration, or its count alone under the sample floor.

    Why: a single signal says what the model believes; its category's record says whether beliefs like it have held up, and a reader deciding what a signal is worth needs both on the same card.

    Recalculated: nothing. The cut is read from the record's cached rows filtered by selection and band; it changes no verdict and no figure.

  4. Effective

    Publication: the Current evidence verdict

    Before: the Overview and Performance opened on a status line over the retrospective record alone, with its bootstrap's workings beside it.

    Now: both open on the Current evidence scoreboard, one of four verdicts, from Insufficient evidence up to Statistically supported edge, composed from the retrospective record and the forward ledger. The retrospective record can hold the verdict down and never raise it: only the forward ledger promotes, from its own floor of settled signals on enough matches, with its mean CLV's interval above nought. The bootstrap's workings moved to Performance's method box and this page's record section.

    Why: a record recut at each row's own evaluation is a retrospective reading, and a profit claim needs a test the rule could not have been fitted to. The verdict says which kind of evidence there is, and how much.

    Recalculated: nothing. The scoreboard prints the record status's figures exactly; the verdict is a reading of them and of the ledger, not a recut.

  5. Effective

    Publication: one vocabulary for every figure

    Before: the section named one figure several ways and several figures one way. A bare edge label stood for the model's expected value, the model's own price and the difference between the two probabilities each went by more than one name, and a count of the rule's rows was called by a word that read as a claim of profit.

    Now: eight terms, each with one entry in the glossary above — Model probability, Market probability, Disagreement (in points), Model price (one over the model probability), Model-implied edge (the model probability × the odds − 1), Signal, Forecast skill and Evidence. A row that meets the rule is a Signal. Green is drawn only where a measured figure's interval lies wholly above nought.

    Why: a reader comparing the model with the market needs each figure to have one name and each name one figure, and a colour that reads as success has to be earned by evidence rather than by a threshold.

    Recalculated: nothing. Every threshold, rule, figure and address is unchanged; only the words and the colours a reader sees changed.

  6. Effective

    One answer per question per match

    Before: a match could carry several signals that answered the same question — the home side to score and the home side to score twice, say — each counted, carded, listed and settled on its own, so one reading of one match could stand in the record as two or three selections.

    Now: every market answers one of five questions — who wins; how many goals the home side scores; how many goals the away side scores; how many goals there are; whether both sides score — and a match gets one answer to each. Among the match's prices that pass every other criterion, the answer to a question with the largest expected value is kept, and every other answer to it says Outranked, stays in All prices with its figures, and is not a signal. Two answers that tie at the stored precision go to the one stored first. One card per match is still presentation and changes no verdict.

    Why: two answers to one question are two stakes on one reading of one match, and a record that counts them twice overstates what it has tested. Which of them is kept is decided by the rule, not by hand.

    Recalculated: every historical selection was re-assessed under this rule at its own evaluation moment. On the 643 settled matches of the day, the record read 1425 selections (465–951–9, -209.4 units, -14.7%) before this entry and 1023 selections (292–725–6, -156.3 units, -15.3%) after it, on the same matches. The two sets are not comparable with each other, and neither is a claim that the model did better or worse for the change.

  7. Effective

    Model vs market: 1X2 published

    Before: the Performance page scored the model on its own, across every market the site supports, with no market side to compare with.

    Now: the page opens with Model vs market: 1X2 — the model's home, draw and away probabilities beside one bookmaker's, margin removed, on the same matches, scored by the multiclass Brier score, the log loss and the average calibration gap, with a base rate beside both scores. The market's price is the last pre-kickoff market price stored at the model's last evaluation; the section is read from 300 paired matches and a band drawn from 30 forecasts. The Overview's calibration block and the Performance hero's calibration tile read the same result.

    Why: a model whose probabilities are set beside prices owes the reader the comparison on equal terms, and the price it is compared with has to be named, with how far before kickoff it was captured.

    Recalculated: nothing. The qualification rule, the record and the Market-specific calibration are unchanged by this entry; the paired comparison is a new reading, not a recut of an old one. A change to the price it is paired with would be an entry of its own here.

  8. Effective

    One rule for every surface

    Before: the section's surfaces each tested their own criteria. The hero counted every positive, unwarned, unstale price whatever the market's measured skill or the outlier line; the cards demanded a flat six hours of freshness; the Market overview and the historical record counted selections on matches with limited data, in markets without measured skill, at prices that were stale at their evaluation, with prices captured after the evaluation that was meant to price them, and with edges of 40% and more. The record was therefore looser than the live board, and the two could disagree about one row.

    Now: one rule (The rule, above) decides for every surface — the hero count, the cards, the Market overview, the league odds lists and the record. At the moment asked about, a price is a signal when the match has not kicked off, the price was captured no later than the evaluation that priced it, it is fresh by the odds schedule's tier for that distance from kickoff, the match carries full data, the model beats its base rate on the market, and the model-implied edge runs from 5% up to but not including 40%. The rule is applied per selection; one card per match is a presentation rule and changes no verdict.

    Why: a record kept by a looser rule than the board it describes is a record of something else. The one rule makes the count on the hero, the card on the board and the row in the record the same statement.

    Recalculated: every historical selection was re-assessed under this rule at its own evaluation moment, and the record on the Overview, the Performance page and the homepage, and each competition's performance section, now print the recalculated figures. They are not comparable with figures published before this date, and neither set is a claim that the model did better or worse for the change.

  9. Effective

    Football model: the non-penalty anchor repaired

    Before: each side's ratings were fitted from the provider's non-penalty expected goals exactly as published. In several competitions that feed is broken — the non-penalty figure about a fifth of the total on matches with no penalty in them — so the model priced every match there at a fraction of the goals the league scores. A side's window was its latest played matches, whether or not the provider had published expected goals for them.

    Now: a match's non-penalty figure is read only when it is believable — when, both sides pooled, it is at least 50% of the match's total expected goals — and the total stands in for it when it is not. A side's window is its latest 20 played matches that carry expected goals. The decay, the shrinkage and the shot-volume blend are unchanged.

    Why: a rating built on a broken feed is wrong by a factor rather than a margin, and in exactly the competitions with nothing else to go on; a result published without expected goals gives the fit nothing to read.

    Recalculated: nothing. Every figure on these pages scores a match by the probability the site published at its evaluation, so a match evaluated before this date was priced by the earlier model and stays as it was published. Figures that mix matches from before and after this date mix two models, and the Model vs market facts strip counts the earlier ones.

For how the probabilities are made in the first place — expected goals, club ratings and the scoreline grid — read How the model works. The comparison itself is on the Betting Overview, and how every signal has done is on Model performance. What the record shows and why is in the dated Research note.