Market-Implied Goals from Betting Odds
Betting odds can be converted into more than home, draw and away probabilities. After the bookmaker margin is removed, the 1X2 market describes the balance between the teams, while the Over/Under market describes the expected scoring environment. A Poisson or Dixon-Coles model can combine those signals and recover market-implied goals: two goal rates, λhome and λaway.
- Convert decimal odds to raw probabilities with 1 / odds.
- Remove the margin from 1X2 and Over/Under separately.
- Under independent Poisson, Over 2.5 mainly determines λhome + λaway.
- 1X2 probabilities mainly determine how the total is divided between the teams; the draw price also carries information about the total itself.
- The markets usually cannot be matched perfectly with only two λ parameters.
- The fitted values are market-implied goal rates, not shot-based xG.
Enter 1X2 and Over/Under odds in the ImpliedScore calculator to estimate market-implied goal rates, the score matrix and market-fit residuals.
Open the score probability calculatorWhat are market-implied goals?
In a football score model, λ is the expected number of goals generated by the probability distribution. If the home team has λ = 1.56, its Poisson distribution has mean 1.56 goals. The away team receives a separate value.
These quantities are not the same as expected goals derived from shots. Shot-based xG assigns values to observed shooting chances during a match, while separate predictive xG models can estimate future chance quality before kickoff. Market-implied λ is recovered from betting prices at a defined market snapshot and reflects whatever information traders and bettors have incorporated into those prices: team strength, injuries, tactics, venue, weather and uncertainty.
This is an inverse problem. A normal score model starts with λ and calculates market probabilities. Here, the market probabilities are known and λ must be found.
ImpliedScore treats 1X2 odds as required input and Over/Under 2.5 as optional but recommended. The extra market is valuable because 1X2 primarily describes who is stronger, while the total market directly constrains how many goals the match is expected to contain. How the fit behaves when the totals market is missing is shown at the end of the worked example.
Market-Implied Goals vs Expected Goals (xG)
| Market-implied λ | Shot-based xG | |
|---|---|---|
| Primary source | Pre-match betting prices | Shot/chance event data |
| Typical timing | Before kickoff or at a quoted market snapshot | During or after the match, or from a separate predictive xG model |
| Meaning | Goal rates that best reproduce de-vigged market probabilities under a chosen score model | Estimated chance quality according to an xG model |
| Model dependence | Depends on de-vig method, Poisson/DC structure, fitting objective and market snapshot | Depends on the xG provider, features, training data and shot definition |
The terms are related but not interchangeable. ImpliedScore does not reconstruct a shot-based expected-goals model from bookmaker prices. It solves an inverse pricing problem and reports market-implied λ: the goal rates that make the chosen score model reproduce the de-vigged market as closely as the fitting objective allows.
Step 1: convert decimal odds into raw probabilities
Decimal odds include the returned stake. The reciprocal is the raw implied probability:
For a 1X2 market quoted at 2.10 / 3.50 / 3.40:
| Outcome | Decimal odds | Raw probability |
|---|---|---|
| Home | 2.10 | 47.62% |
| Draw | 3.50 | 28.57% |
| Away | 3.40 | 29.41% |
| Total | 105.60% |
The total exceeds 100% because these are prices, not a fair probability distribution. The excess above 100% is the overround. For one-price conversion, see the Implied Probability Calculator; for complete-market de-vigging, see the No-Vig Calculator.
Step 2: remove the bookmaker margin
ImpliedScore uses proportional normalization as its reproducible production baseline. Divide each reciprocal price by the sum of all reciprocal prices in that market:
The 1X2 example becomes:
| Outcome | Raw probability | Fair probability | Fair odds |
|---|---|---|---|
| Home | 47.62% | 45.09% | 2.22 |
| Draw | 28.57% | 27.06% | 3.70 |
| Away | 29.41% | 27.85% | 3.59 |
| Total | 105.60% | 100.00% |
Remove the margin from each market independently. Do not combine the three 1X2 prices and the two Over/Under prices into one five-outcome normalization: 1X2 and O/U are separate markets whose outcomes overlap rather than form one mutually exclusive set.
Over/Under example
For Over 2.5 at 1.85 and Under 2.5 at 2.00:
| Outcome | Odds | Raw probability | Fair probability |
|---|---|---|---|
| Over 2.5 | 1.85 | 54.05% | 51.95% |
| Under 2.5 | 2.00 | 50.00% | 48.05% |
| Total | 104.05% | 100.00% |
De-vigging is a modelling choice. Proportional normalization is the current ImpliedScore baseline. Alternative margin-removal approaches are discussed by Štrumbelj (2014) and Clarke, Kovalchik & Ingram (2017). The public No-Vig Calculator exposes alternative methods as sensitivity comparisons rather than silently changing historical tracker outputs.
Step 3: use Over/Under to infer the total goal rate
Under an independent Poisson model:
The sum of two independent Poisson variables is also Poisson:
Over 2.5 means at least three total goals. Its model probability is:
The fair Over 2.5 probability can therefore be converted into a total goal rate by solving:
There is no elementary closed-form solution for Λ, but the left side is monotonic for positive Λ. A one-dimensional root finder such as bisection or Brent’s method solves it reliably.
A fair Over 2.5 probability of 51.95% implies Λ ≈ 2.75 under independent Poisson. That is the total scoring anchor before the 1X2 information is used to divide the goals between the teams.
Other total lines
The same idea applies to any half-goal line. For Over n + 0.5:
Multiple total lines can provide additional constraints, but they also expose model inconsistency: one Poisson Λ may not reproduce every quoted Over/Under line exactly.
Step 4: use 1X2 to split the total between the teams
Over/Under constrains the sum of the rates. It does not say which team is expected to score more. The 1X2 market supplies that information.
If total Λ is fixed, write the home share as s:
For each candidate value of s, construct the score matrix and sum the cells:
Increasing the home share raises home-win probability and lowers away-win probability. The draw probability is generally highest when the rates are similar and falls as the match becomes more imbalanced.
The Skellam representation
The goal difference Z = X − Y follows a Skellam distribution under independent Poisson scoring:
Here, I is the modified Bessel function of the first kind. Then: home win is P(Z > 0), draw is P(Z = 0), and away win is P(Z < 0). In implementation, direct matrix summation is often simpler and also produces all exact-score probabilities needed later.
Why the market usually cannot be matched exactly
The model has two free rates: λhome and λaway. The target data contain:
- home, draw and away probabilities, with two independent degrees of freedom because they sum to 1;
- one independent Over/Under probability because Over and Under also sum to 1.
That gives three independent market constraints for two model parameters. The fitting objective may contain four residual terms—home, draw, away and Over—but the first three include one linear redundancy.
An exact solution exists only when the fair market probabilities happen to lie on the two-dimensional probability surface generated by the chosen score model. In real data, they usually do not.
Sources of inconsistency
- The true score distribution is not exactly independent Poisson.
- The 1X2 and totals markets may have been captured at different times.
- Prices may come from different bookmakers with different opinions and margins.
- Displayed odds are rounded.
- Proportional de-vigging may not recover the bookmaker’s internal fair probabilities.
- Market-specific margin and liability management distort prices differently.
- Dixon-Coles dependence changes low-score and draw probabilities.
Residuals are information, not merely error. A large draw residual can indicate that the market requires more low-score dependence than independent Poisson allows. A large totals residual can indicate conflicting inputs or a poor fit to the chosen goal model.
Step 5: choose λ by minimizing market-fit error
ImpliedScore uses a weighted squared-error objective:
The Over 2.5 residual receives double weight because it is the direct anchor for total goals. The optimizer searches for the positive λ pair with the lowest error.
A comparable joint squared-error calibration to 1X2 and Over/Under markets is used by Clegg, Song & Cartlidge (2026), who apply it to a Weibull accelerated-failure-time goal model fitted to Betfair Exchange prices rather than to an independent-Poisson score matrix.
Why not solve only two equations?
It is possible to solve the Over probability exactly for Λ and then match one 1X2 quantity, such as home-win probability, by adjusting the split. But the resulting draw and away probabilities can remain materially wrong.
Least squares uses all available information and exposes the compromise. The Model Fit output should display the residual for each target rather than presenting λ as an exact market identity.
Weights are part of the model
A weight of 2 for Over 2.5 is not a universal statistical law and not an estimated reliability ratio for the totals market. It expresses a design priority. The three 1X2 residuals are also not independent observations: they contain one exact linear redundancy, so the objective effectively balances two 1X2 contrasts against one double-weighted totals term. Other implementations may:
- use equal weights;
- weight by estimated market reliability;
- divide by forecast variance;
- fit log-odds rather than probability differences;
- use a likelihood or proper scoring objective;
- force one market to match exactly and minimize the others.
Results from different methods should not be compared without documenting the objective, de-vig method, parameter bounds and score-grid truncation.
Worked example: from five prices to two λ values
Use the following market:
| Home | Draw | Away | Over 2.5 | Under 2.5 |
|---|---|---|---|---|
| 2.10 | 3.50 | 3.40 | 1.85 | 2.00 |
1. De-vigged targets
| Target | Fair probability |
|---|---|
| Home | 45.09% |
| Draw | 27.06% |
| Away | 27.85% |
| Over 2.5 | 51.95% |
2. Independent-Poisson fit
Minimizing the stated objective gives approximately:
The fitted total is 2.74 goals, slightly below the 2.75 implied by the O/U market alone because the optimizer also tries to reproduce 1X2.
A continuous least-squares reference calculation gives λ home = 1.55571695 and λ away = 1.18476574, with objective E ≈ 0.0006051630. The table rounds only the final probabilities and residuals. A finite production grid can return slightly different rates; this reference calculation does not test the installed solver.
| Market | Fair target | Poisson model | Residual |
|---|---|---|---|
| Home | 45.09% | 45.90% | +0.81pp |
| Draw | 27.06% | 25.09% | −1.96pp |
| Away | 27.85% | 29.01% | +1.16pp |
| Over 2.5 | 51.95% | 51.62% | −0.32pp |
The largest mismatch is the draw. Independent Poisson assigns too little probability to draws for this market combination and compensates by raising both home and away wins.
3. Dixon-Coles fit with ρ = −0.03
If the low-score correction is included during optimization, the fitted values become approximately:
The continuous reference rates are λ home = 1.55930700 and λ away = 1.18553647, with E ≈ 0.0002582420. The fitted rates move by less than 0.01 goals; most of the improvement comes from τ redistributing probability toward 0-0 and 1-1.
| Market | Fair target | Dixon-Coles model | Residual |
|---|---|---|---|
| Home | 45.09% | 45.61% | +0.52pp |
| Draw | 27.06% | 25.78% | −1.28pp |
| Away | 27.85% | 28.61% | +0.76pp |
| Over 2.5 | 51.95% | 51.73% | −0.22pp |
A negative ρ increases 0-0 and 1-1 while decreasing 1-0 and 0-1. That raises total draw probability and improves this example’s fit, but it still does not reproduce all targets exactly.
4. What if Over/Under odds are unavailable?
With 1X2 only, the Over residual is dropped and the objective becomes:
The three fair probabilities sum to one, so they carry two independent degrees of freedom — the same number of constraints as two goal rates. This count alone does not guarantee a feasible or unique solution within a chosen model and search bounds. With ρ fixed at zero in this example, the draw price carries most of the information about the total, while the Home–Away imbalance sets the split. For the same market, the 1X2-only fit gives:
The continuous reference fit uses λ home = 1.38310054 and λ away = 1.02512718. Its three 1X2 residuals are numerically zero: two parameters can reproduce two degrees of freedom exactly whenever the fair probabilities lie on the model surface, as they do here. A perfect Model Fit readout in this mode reflects exact identification, not validation — there is nothing left to check the fit against.
The cost is anchoring. The fitted total drops to 2.41 goals and implies an Over 2.5 probability of about 43.2%, far from the 51.9% quoted by the totals market: the two markets disagree by roughly a third of a goal here, and without Over/Under the entire total rests on the draw price alone. The estimate is well-defined but less strongly anchored than a fit that also uses a totals market.
Enter λhome = 1.56 and λaway = 1.18 in the Poisson Football Calculator to see the full independent-Poisson score matrix behind this fit.
Open the Poisson Football CalculatorStep 6: turn λ into a score probability matrix
Once λ values are fixed, calculate every exact-score probability. Under independent Poisson:
For the Dixon-Coles example above, the largest cells are approximately:
| Score | Probability | Fair odds |
|---|---|---|
| 1-1 | 12.24% | 8.17 |
| 1-0 | 9.66% | 10.35 |
| 2-1 | 9.26% | 10.80 |
| 2-0 | 7.81% | 12.80 |
| 0-1 | 7.26% | 13.77 |
The score matrix is then summed to calculate:
- home, draw and away fair odds;
- Over/Under 1.5, 2.5, 3.5 and other totals;
- both teams to score;
- team-total probabilities and fair odds;
- double chance and draw-no-bet;
- Asian Handicap settlement probabilities and fair odds;
- the most likely scorelines.
The most likely score is still unlikely. Here, 1-1 ranks first but has only about a 12.2% probability. The output is a distribution over many possible scores, not a deterministic correct-score prediction.
The table above uses the unrounded fixed-ρ reference rates and Dixon-Coles correction (ρ = −0.03). The Poisson Football Calculator linked above uses ρ = 0, so the same rates give slightly different low-score cells there: 1-1 is about 11.9% instead of 12.2%. To start from projected goal rates and choose ρ directly, use the xG to Fair Odds Calculator.
How Dixon-Coles changes the inverse fit
Dixon-Coles multiplies four low-score cells by a correction function τ:
| Score | τ multiplier |
|---|---|
| 0-0 | 1 − λhλaρ |
| 1-0 | 1 + λaρ |
| 0-1 | 1 + λhρ |
| 1-1 | 1 − ρ |
A “valid ρ” is not a fixed interval. For the calculator and fitting loop, all four multipliers are required to remain strictly positive. This bounds ρ by the current goal rates:
Because the λ values move during optimization, the bound moves with them: every candidate parameter set must be checked, and any set that makes a τ multiplier or a score probability negative must be rejected. For the example above, ρ = −0.03 sits far inside the valid range (roughly −0.64 to +0.54 at the fitted rates); violations appear with strongly negative corrections at high goal rates, where the lower bound tightens toward zero.
If the final matrix will use Dixon-Coles, the correction should also be active while λ is fitted. Fitting independent Poisson rates first and modifying four cells afterward produces outputs that no longer minimize the stated market-fit objective.
At fixed λ, the τ adjustment preserves total probability and both marginal goal distributions. It also leaves Over 2.5 unchanged because all four corrected cells are Under 2.5 and their net probability change is zero. Lower total lines are different: O/U 0.5 and O/U 1.5 can change directly. The optimal λ values can also shift when Dixon-Coles is included in the fitting loop because 1X2 residuals change.
See the Dixon-Coles model guide for the derivation, interpretation of ρ and its limitations.
Implementation workflow
- Validate the inputs. Require odds above 1.00, reject missing 1X2 outcomes and ensure Over/Under prices refer to the same line and match state.
- Convert prices to raw probabilities. Calculate 1 / odds for every selection.
- De-vig each market independently. Normalize 1X2 to 100%, then normalize Over/Under to 100%.
- Choose the score model. Use independent Poisson or set a Dixon-Coles ρ, rejecting during the search any candidate λ pair that would make a τ multiplier or score probability negative.
- Search the λ space. Evaluate positive λ pairs, calculate the matrix and minimize the documented objective.
- Refine around the best pair. Use a finer grid or a bounded numerical optimizer.
- Check probability mass. If the matrix is truncated, report the omitted tail and expand the grid when necessary.
- Publish residuals. Show target versus model for Home, Draw, Away and Over 2.5.
- Generate derived markets. Sum the full computational grid, not only the visible heatmap.
Minimal pseudocode
Grid search versus continuous optimization
| Method | Advantages | Risks |
|---|---|---|
| Coarse-to-fine grid | Transparent, deterministic, easy to debug | Resolution depends on step size; computationally heavier |
| Bounded local optimizer | Fast and precise near a smooth minimum | Can depend on initial values and implementation details |
| Global optimizer | More robust to irregular objectives | Usually unnecessary for a well-bounded two-parameter problem |
| Two-stage total/split solution | Fast and interpretable | May privilege O/U and fail to minimize all residuals jointly |
ImpliedScore uses a two-pass grid search: a coarse pass from 0.20 to 6.00 goals per team, followed by a finer pass within ±0.08 of the best pair. A solution at or near either search boundary should be treated as a warning that the input market lies outside the normal football range or that the bounds need to be expanded. The full implementation is documented in the ImpliedScore methodology.
Use the main calculator linked above to enter market odds and inspect fitted λ, score probabilities and market-fit residuals together.
Limitations and interpretation
λ is model-dependent
The same odds can produce different λ values under independent Poisson, Dixon-Coles, bivariate Poisson or another score model. There is no model-free transformation from five prices to two team goal rates.
λ is de-vig-dependent
Proportional, Shin and power methods produce different fair probabilities. The difference may be small in balanced, low-margin markets and larger for strong favourites or high-margin books.
Market timing matters
Opening odds, current odds and closing odds represent different information sets. Mixing a closing 1X2 market with an earlier total market creates artificial inconsistency.
Two parameters cannot encode every market
Correct score, Asian handicap, team totals and BTTS may all contain information that a two-rate Poisson model cannot reproduce simultaneously.
Constant scoring rates are an approximation
Poisson treats each team’s goal rate as fixed throughout the match. Actual rates change with time, score state, substitutions and red cards.
Market-implied does not mean objectively true
The method translates prices into a coherent score distribution. It does not prove that the bookmaker or market probabilities are accurate, nor does it create an independent betting edge.
A low fitting error is not validation
Good reproduction of the input markets shows internal consistency. Predictive quality must be evaluated against future results with calibration, Brier score, log loss and out-of-sample testing. See football probability calibration.
Practical checklist
- Use odds from the same bookmaker and timestamp where possible.
- Remove margin separately for 1X2 and Over/Under.
- Document the de-vig method.
- Use O/U to anchor total goals and 1X2 to determine team balance.
- Include Dixon-Coles during fitting if it will be used in the final matrix.
- Show residuals rather than claiming an exact conversion.
- Check tail probability outside the computational grid.
- Call the outputs market-implied goals (λ), not observed or shot-based xG.
References
- Maher, M. J. (1982). Modelling Association Football Scores . Statistica Neerlandica, 36(3), 109–118.
- Dixon, M. J. & Coles, S. G. (1997). Modelling Association Football Scores and Inefficiencies in the Football Betting Market . Journal of the Royal Statistical Society: Series C, 46(2), 265–280.
- Štrumbelj, E. (2014). On Determining Probability Forecasts from Betting Odds . International Journal of Forecasting, 30(4), 934–943.
- Clarke, S. R., Kovalchik, S. & Ingram, M. (2017). Adjusting Bookmaker’s Odds to Allow for Overround . American Journal of Sports Science, 5(6), 45–49.
- Clegg, L., Song, Z. & Cartlidge, J. (2026). A Market-Calibrated Accelerated Failure Time Model for In-Play Football Forecasting . arXiv preprint arXiv:2605.16066.