Match model — probabilities and prices

Best value today across all competitions, ranked by expected value rather than by how likely they are to win

Stake is blank unless a selection clears every gate: positive expected value, at least 5 percentage points of edge over the de-vigged line, and not more than 25 (a disagreement that large is usually the model missing something, not an opportunity). Most rows will be blank most days. Selections marked longshot are below 10% — their EV figure divides by a very small number and should not be trusted. thin data means one side has too few matches in the model's history for its rating to mean anything; the huge edge shown is the model's ignorance, not the market's mistake.

Best goal markets over/under and both teams to score, ranked by expected value

Kept separate from the match-winner table because they are a different bet: independent of who wins, driven by the total-goals distribution rather than the difference between the two sides. Worth a look for two reasons — a goal model is being asked the question it was actually built to answer, and a totals line is far less correlated with the rest of a slip than another favourite. Against that, books charge more margin here than on the match winner, so a given edge has further to travel.

No goal markets priced. These need over/under and both-teams-to-score odds, which only arrive with --odds.

Same-match combinations priced from the joint distribution, so correlation is handled correctly

The vs naive column is how far a bookmaker would be off if they priced the combination by multiplying the two legs together. A large positive number means the events happen together more often than independence implies — draws and low-scoring games, for instance — so a multiplied price is too generous.

Long builders 10+ legs, 1,000+ fair odds, reusing the same matches

Legs from the same match are combined by intersecting the score matrix, so their correlation is exact. Legs from different matches are multiplied, which is correct because they are independent. Most bookmakers' builders multiply everything, and on correlated legs that is wrong by a factor shown against each slip below — typically 2 to 3x here.
Slips are ranked by true probability, not by that factor. At a fixed odds target the most mispriced slip is simply the one stacking the most redundant legs, which is a worse bet dressed up; the least unlikely route to the target is the useful one. Leg groups that collapse to a single selection are excluded, and any that survive are flagged.

Corners a different distribution and a different estimator

Corners are not Poisson - variance over mean is 1.17 on the total against 0.99 for goals - so these come from a negative binomial. Using the goal machinery would run about two points light on over 13.5, always in the same direction. Reusing the Dixon-Coles team ratings was tried too and finished worse than predicting the league average; these rates are time-weighted rolling means instead.
Out of sample on 27,773 matches the totals beat the base rate by 0.0108 of log loss and most corners by 0.0407, at 58.0% accuracy against 53.6%. Most corners is the stronger market by some distance - corner dominance is a stable team property, the number of corners in a match much less so. No bookmaker corner prices were fetched, so these are model numbers with no edge attached.

Fixtures where either side has fewer than 10 matches of corner history are absent rather than averaged, and four competitions in this dataset carry no corner data at all. Corner totals and corner results are never combined with each other: they come from separate estimators with no joint distribution between them.

Odd / even total goals offered, not recommended

Across the 16 fixtures on this page the model puts odd between 48.2% and 49.3% - a spread of 1.1 points. That is not a limitation of this model, it is the arithmetic: for a total with mean L the chance of an odd number is (1 - e^-2L) / 2, which is 49.1% at a 2.0-goal match and 50.0% at 3.5. Measured on 65,170 real matches the split is 49.90 / 50.10.
A bookmaker's margin on this market is several times the whole range the model can express, so there is no price at which the numbers below constitute an edge. They are here for completeness.

Over/under builder pick a side and a line, or let the model choose

Out of fold during training this market scored 0.6103 against 0.6920 for always predicting the base rate, over 97,770 observations. That is the reason this builder exists for totals and not for every market on the page.

Both sides of every line are listed, so choosing the side the model dislikes is allowed — it just gives you the complement, since Over 2.5 and Under 2.5 come from one distribution and sum to exactly one. Model's choice picks the line it is most confident about per fixture, marked best, which is rarely the 2.5 every coupon leads with: 2.5 sits nearest the median total, which is exactly where a goal model has least to say.

Whatever you choose, the slip takes one leg per fixture. Over 0.5, 1.5 and 2.5 on the same match are not three bets — they nearly all win or lose together, and multiplying them would state a confidence the model does not have. Confidence is the calibrated model probability, not a price: a 4-leg slip at 80% a leg lands 41% of the time, and the margin still compounds at roughly 5% per leg.

Six-leg builder every leg above 50%, drawn from at least three different markets

Legs come from different matches, so the probabilities do multiply — that part is honest. Margin cost is the share of fair value lost to compounding vig at 5% per leg. Weakest leg is the one carrying the slip: a chain is only as good as that number, and it is the first thing a bookmaker chooses when they build these for you.

Accumulators shown with what they cost you

Legs from different matches are independent, so the probabilities multiply — and so does the bookmaker's margin. The final column is the share of fair value lost to compounding vig at 5% per leg, before your model is even wrong about anything. Accumulators are on this page because people want them, not because they are a good idea.

By competition 16 fixtures across 12 competitions · 0/48 selections priced

Stake is blank unless a selection clears every gate: positive expected value, at least 5 percentage points of edge over the de-vigged line, and not more than 25 (a disagreement that large is usually the model missing something, not an opportunity). Most rows will be blank most days. Selections marked longshot are below 10% — their EV figure divides by a very small number and should not be trusted. thin data means one side has too few matches in the model's history for its rating to mean anything; the huge edge shown is the model's ignorance, not the market's mistake.

Stake is blank unless a selection clears every gate: positive expected value, at least 5 percentage points of edge over the de-vigged line, and not more than 25 (a disagreement that large is usually the model missing something, not an opportunity). Most rows will be blank most days. Selections marked longshot are below 10% — their EV figure divides by a very small number and should not be trusted. thin data means one side has too few matches in the model's history for its rating to mean anything; the huge edge shown is the model's ignorance, not the market's mistake.

Stake is blank unless a selection clears every gate: positive expected value, at least 5 percentage points of edge over the de-vigged line, and not more than 25 (a disagreement that large is usually the model missing something, not an opportunity). Most rows will be blank most days. Selections marked longshot are below 10% — their EV figure divides by a very small number and should not be trusted. thin data means one side has too few matches in the model's history for its rating to mean anything; the huge edge shown is the model's ignorance, not the market's mistake.

Stake is blank unless a selection clears every gate: positive expected value, at least 5 percentage points of edge over the de-vigged line, and not more than 25 (a disagreement that large is usually the model missing something, not an opportunity). Most rows will be blank most days. Selections marked longshot are below 10% — their EV figure divides by a very small number and should not be trusted. thin data means one side has too few matches in the model's history for its rating to mean anything; the huge edge shown is the model's ignorance, not the market's mistake.

Stake is blank unless a selection clears every gate: positive expected value, at least 5 percentage points of edge over the de-vigged line, and not more than 25 (a disagreement that large is usually the model missing something, not an opportunity). Most rows will be blank most days. Selections marked longshot are below 10% — their EV figure divides by a very small number and should not be trusted. thin data means one side has too few matches in the model's history for its rating to mean anything; the huge edge shown is the model's ignorance, not the market's mistake.

Stake is blank unless a selection clears every gate: positive expected value, at least 5 percentage points of edge over the de-vigged line, and not more than 25 (a disagreement that large is usually the model missing something, not an opportunity). Most rows will be blank most days. Selections marked longshot are below 10% — their EV figure divides by a very small number and should not be trusted. thin data means one side has too few matches in the model's history for its rating to mean anything; the huge edge shown is the model's ignorance, not the market's mistake.

Stake is blank unless a selection clears every gate: positive expected value, at least 5 percentage points of edge over the de-vigged line, and not more than 25 (a disagreement that large is usually the model missing something, not an opportunity). Most rows will be blank most days. Selections marked longshot are below 10% — their EV figure divides by a very small number and should not be trusted. thin data means one side has too few matches in the model's history for its rating to mean anything; the huge edge shown is the model's ignorance, not the market's mistake.

Stake is blank unless a selection clears every gate: positive expected value, at least 5 percentage points of edge over the de-vigged line, and not more than 25 (a disagreement that large is usually the model missing something, not an opportunity). Most rows will be blank most days. Selections marked longshot are below 10% — their EV figure divides by a very small number and should not be trusted. thin data means one side has too few matches in the model's history for its rating to mean anything; the huge edge shown is the model's ignorance, not the market's mistake.

Stake is blank unless a selection clears every gate: positive expected value, at least 5 percentage points of edge over the de-vigged line, and not more than 25 (a disagreement that large is usually the model missing something, not an opportunity). Most rows will be blank most days. Selections marked longshot are below 10% — their EV figure divides by a very small number and should not be trusted. thin data means one side has too few matches in the model's history for its rating to mean anything; the huge edge shown is the model's ignorance, not the market's mistake.

Stake is blank unless a selection clears every gate: positive expected value, at least 5 percentage points of edge over the de-vigged line, and not more than 25 (a disagreement that large is usually the model missing something, not an opportunity). Most rows will be blank most days. Selections marked longshot are below 10% — their EV figure divides by a very small number and should not be trusted. thin data means one side has too few matches in the model's history for its rating to mean anything; the huge edge shown is the model's ignorance, not the market's mistake.

Stake is blank unless a selection clears every gate: positive expected value, at least 5 percentage points of edge over the de-vigged line, and not more than 25 (a disagreement that large is usually the model missing something, not an opportunity). Most rows will be blank most days. Selections marked longshot are below 10% — their EV figure divides by a very small number and should not be trusted. thin data means one side has too few matches in the model's history for its rating to mean anything; the huge edge shown is the model's ignorance, not the market's mistake.

Stake is blank unless a selection clears every gate: positive expected value, at least 5 percentage points of edge over the de-vigged line, and not more than 25 (a disagreement that large is usually the model missing something, not an opportunity). Most rows will be blank most days. Selections marked longshot are below 10% — their EV figure divides by a very small number and should not be trusted. thin data means one side has too few matches in the model's history for its rating to mean anything; the huge edge shown is the model's ignorance, not the market's mistake.