xG to Fair Odds Calculator: Convert Expected Goals into Probabilities
Two expected-goal inputs can be converted into a complete probability model for a football match. Treat the home and away goal expectations as scoring rates, build a Poisson or Dixon-Coles score matrix, sum the relevant cells for 1X2, totals and BTTS, then invert each probability to obtain model-implied fair odds.
- Use pre-match projected goal rates, not the final xG from a match that has already been played.
- Set λhome and λaway from the two expected-goal estimates.
- Calculate each score with the Poisson formula and combine the two team distributions.
- Sum score cells to obtain 1X2, Over/Under, BTTS and team-total probabilities.
- Convert probability to fair decimal odds with 1 ÷ probability.
- Fair odds are conditional on the xG forecast and score model; they are not guaranteed true prices.
xG to fair odds calculator
What does expected goals mean?
A shot-level expected goals model assigns each attempt a probability between 0 and 1 of becoming a goal. The estimate can use location, angle, body part, assist type, defensive pressure, goalkeeper position and other event or tracking features.
If a team takes shots with xG values of 0.10, 0.25 and 0.40, the cumulative xG is:
This means the expected number of goals from that specific set of chances is 0.75 under the shot model. It does not mean that the team had a 75% probability of scoring at least once, nor that 0.75 can be inserted automatically into a pre-match betting model.
xG is an expected value. An expected value summarizes the centre of a probability distribution. It does not uniquely determine the full distribution.
Three different quantities are often called “xG”
| Quantity | When it is known | What it measures | Can it be used directly as pre-match λ? |
|---|---|---|---|
| Shot-level cumulative xG | During or after a match | Expected goals from the chances that actually occurred | No; the future shot set was unknown before kickoff |
| Historical average xG | Before the next match | Past chance creation or concession over a chosen sample | Not without opponent, venue, recency and lineup adjustments |
| Pre-match projected xG | Before kickoff | A forecast of each team’s expected goals in the upcoming fixture | Yes, after calibration to the intended score model |
| Market-implied λ | Before kickoff | Goal rates recovered from margin-free betting probabilities | Yes, but it is derived from odds rather than shot data |
This page uses the term projected xG for the pre-match home and away goal expectations supplied to the score model.
Market-implied goal rates solve a different problem. They begin with 1X2 and Over/Under odds and work backwards to λ. That inverse process is explained in Market-Implied Goals from Betting Odds.
Hindsight error: using the final xG totals from a completed match to produce “pre-match fair odds” is not forecasting. It answers a retrospective question: what result probabilities correspond to a Poisson/DC approximation with those realized xG totals? For probabilities conditional on the individual shots themselves, retain the shot-level probabilities and an explicit dependence assumption rather than only their sum.
Why cumulative xG does not uniquely determine goal probabilities
Suppose the probabilities of scoring from a known set of shots are p1, p2, …, pn. If the shot outcomes are treated as independent Bernoulli variables, the number of goals follows a Poisson-binomial distribution, not necessarily a Poisson distribution.
Linearity of expectation does not require independence: E[G] = Σpi holds under any dependence between shots. Independence is needed for the goal count to follow the Poisson-binomial distribution and for product formulas such as P(G = 0) = Π(1 − pi).
The probability of scoring no goals from that known shot set is:
Therefore, the probability of scoring at least once is:
Same xG, different scoring probability
| Chance pattern | Total xG | P(at least one goal) | P(exactly two goals) |
|---|---|---|---|
| One chance at 0.80 | 0.80 | 80.00% | 0.00% |
| Eight chances at 0.10 | 0.80 | 56.95% | 14.88% |
| Poisson approximation with λ = 0.80 | 0.80 | 55.07% | 14.38% |
Both shot sets have the same expected value, but their distributions differ. A Poisson approximation is much closer to many small-probability shots than to one very large chance.
Why Poisson is still useful pre-match
Before kickoff, the exact future shot probabilities are not known. A forecast usually produces only a team-level goal expectation. Poisson supplies a transparent one-parameter count model from that single mean, making it practical for a pre-match score matrix.
The approximation should be tested rather than assumed perfect. Penalties, one-on-ones, rebounds, correlated attacks and changing game state can all make the real goal distribution differ from a fixed-rate Poisson model.
Step 1: treat projected xG as the Poisson goal rate
Let:
Under independent Poisson scoring, the probability that a team with rate λ scores exactly k goals is:
The expected value and variance both equal λ. If the home team is projected at 1.70 xG:
| Home goals | Probability with λ = 1.70 |
|---|---|
| 0 | 18.27% |
| 1 | 31.06% |
| 2 | 26.40% |
| 3 | 14.96% |
| 4 or more | 9.32% |
Repeat the same calculation with λaway. The two marginal distributions are then combined into a joint score matrix.
For the underlying count model and its assumptions, see Poisson Football Calculator and model guide.
Step 2: build the exact-score matrix
If home and away goals are independent conditional on their rates, the joint probability of home score i and away score j is:
Each cell of the matrix is one exact score. For λhome = 1.70 and λaway = 1.10:
| Home \ Away | 0 | 1 | 2 | 3 | 4 |
|---|---|---|---|---|---|
| 0 | 6.08% | 6.69% | 3.68% | 1.35% | 0.37% |
| 1 | 10.34% | 11.37% | 6.25% | 2.29% | 0.63% |
| 2 | 8.79% | 9.67% | 5.32% | 1.95% | 0.54% |
| 3 | 4.98% | 5.48% | 3.01% | 1.10% | 0.30% |
| 4 | 2.12% | 2.33% | 1.28% | 0.47% | 0.13% |
The displayed 0–4 grid omits higher scores. Calculations should use a larger internal grid and verify that the omitted tail probability is negligible.
A score matrix contains the complete model. 1X2, totals, BTTS, team totals and correct-score prices are different sums of the same cells. They should not be calculated with unrelated formulas that can contradict one another.
Step 3: calculate football market probabilities
Home, draw and away
Over and Under 2.5
The total of two independent Poisson variables is Poisson with rate Λ = λhome + λaway. Therefore:
The matrix method gives the same result by summing all cells where i + j ≥ 3.
Both teams to score
Under independence:
Equivalently:
Team to score
Other derived markets
- Over/Under 0.5, 1.5, 3.5 and higher lines;
- home and away team totals;
- double chance;
- draw no bet;
- win to nil;
- exact total goals;
- correct-score groups.
Asian quarter-goal lines require settlement-aware expected-return calculations because half the stake can be assigned to each adjacent half- or whole-goal line. They should not be priced by treating the quoted line as one simple binary event.
Step 4: convert probabilities into fair odds
For markets without pushes or partial settlement — 1X2, Over/Under 2.5, BTTS, team to score, exact scores and half-goal handicap lines — fair decimal odds equal the reciprocal of the model probability:
Markets that can return all or part of the stake must be priced from settlement-weighted expected return instead. For a win–push–loss market such as draw no bet, a whole team total or a whole Asian handicap:
For Asian quarter lines with half wins and half losses:
Example: home draw no bet with the worked-example rates has win 51.40%, push 24.01% and loss 24.58%, giving fair odds 1 + 0.2458/0.5140 = 1.478 — not 1/0.5140 = 1.95. The 1X2 and totals examples below are push-free, so the simple reciprocal applies.
A 51.40% home-win probability corresponds to:
These are model-implied fair odds: the break-even price under the supplied probability estimate, before commission, tax, limits, execution uncertainty or model error.
Comparing fair odds with offered odds
If decimal odds O are available and the model probability is p, the expected return per unit staked is:
With p = 0.5140 and offered odds 2.05:
This is only as reliable as the probability model. A positive mathematical expectation calculated from miscalibrated xG inputs is not a real edge. Model uncertainty is usually much larger than the displayed rounding suggests.
Worked example: home xG 1.70, away xG 1.10
Assume the two values are calibrated pre-match goal expectations:
Independent-Poisson probabilities and fair odds
| Market | Probability | Fair decimal odds |
|---|---|---|
| Home win | 51.40% | 1.95 |
| Draw | 24.01% | 4.16 |
| Away win | 24.58% | 4.07 |
| Over 2.5 | 53.05% | 1.88 |
| Under 2.5 | 46.95% | 2.13 |
| BTTS Yes | 54.53% | 1.83 |
| BTTS No | 45.47% | 2.20 |
| Home team to score | 81.73% | 1.22 |
| Away team to score | 66.71% | 1.50 |
Most likely exact scores
| Rank | Score | Probability | Fair odds |
|---|---|---|---|
| 1 | 1-1 | 11.37% | 8.79 |
| 2 | 1-0 | 10.34% | 9.67 |
| 3 | 2-1 | 9.67% | 10.35 |
| 4 | 2-0 | 8.79% | 11.38 |
| 5 | 0-1 | 6.69% | 14.95 |
“Most likely” does not mean likely in absolute terms. The top score, 1-1, still has only an 11.37% probability. Almost 89% of the distribution lies elsewhere.
ImpliedScore’s public calculator derives λ from market odds rather than accepting xG directly, but the same score-matrix engine converts the fitted rates into these outputs.
Open the score probability calculatorHow Dixon-Coles changes xG-based fair odds
Independent Poisson can misrepresent local dependence around 0-0, 1-0, 0-1 and 1-1. Dixon-Coles multiplies those four cells by a τ correction controlled by ρ.
For the calculator and fitting loop, all four τ multipliers must remain strictly positive: max(−1/λhome, −1/λaway) < ρ < min(1/(λhomeλaway), 1). For the worked-example rates the strict interval is approximately −0.588 < ρ < 0.535.
With λhome = 1.70, λaway = 1.10 and ρ = −0.03, each affected cell moves by approximately 0.341 percentage points:
| Score | Independent Poisson | Dixon-Coles | Change |
|---|---|---|---|
| 0-0 | 6.081% | 6.422% | +0.341pp |
| 1-0 | 10.338% | 9.997% | −0.341pp |
| 0-1 | 6.689% | 6.348% | −0.341pp |
| 1-1 | 11.371% | 11.713% | +0.341pp |
| Market | Independent Poisson | Dixon-Coles ρ = −0.03 |
|---|---|---|
| Home win | 51.40% | 51.06% |
| Draw | 24.01% | 24.69% |
| Away win | 24.58% | 24.24% |
| Over 2.5 | 53.05% | 53.05% |
| BTTS Yes | 54.53% | 54.87% |
Over 2.5 remains unchanged at fixed λ because all four adjusted scorelines are Under 2.5 and their net probability movement is zero. Lower total lines are different: O/U 0.5 and O/U 1.5 can change directly. Draw and BTTS also change because the correction reallocates probability within the low-score region.
ρ should be estimated or externally validated. It is not a universal football constant. See the Dixon-Coles model explained.
How to obtain pre-match projected xG
Converting λ to probabilities is mechanical. Producing reliable λ values is the harder forecasting problem.
A basic attack–defence structure
A simple model can estimate a team’s attacking strength and defensive weakness relative to the league, then combine them with venue:
This is the xG analogue of the attack–defence structure used in historical Poisson models. See Maher’s football Poisson model.
Inputs that may require adjustment
- home and away splits;
- opponent strength;
- recency and changing team quality;
- promoted or relegated teams;
- expected lineups and goalkeeper changes;
- penalty allocation;
- competition and match state;
- rest, travel and schedule congestion;
- provider-specific xG definitions.
Why a raw rolling average is weak
Averaging the last five xG values gives every opponent and match context the same interpretation. A team can post high xG because it faced weak defences, trailed early and attacked aggressively, or received several penalties. Small windows are volatile; long windows adapt slowly.
A stronger workflow estimates latent attack and defence parameters with recency weights or a dynamic model, then calibrates the resulting λ values on future matches.
Do not mix providers silently
Different xG systems use different features, training data, penalty values and treatment of blocked shots or goalkeeper information. Two providers can assign different xG to the same attempt. A model trained on one source should not be updated with another source without revalidation.
Projected xG needs its own model. “Team A averaged 1.7 xG, therefore λ = 1.7 next week” is an assumption, not a validated forecast.
How to validate xG-derived probabilities
Validate the entire pipeline on matches that were not used to estimate or tune the projected xG model.
- Freeze the pre-match inputs. Store λhome, λaway, timestamps, lineups available at that time and the model version.
- Use chronological testing. Train on earlier matches and score later matches to avoid future information leaking into attack or defence ratings.
- Test goal-count calibration. Compare predicted and observed frequencies for 0, 1, 2, 3 and higher goals.
- Test market calibration. Use classwise reliability diagrams for Home, Draw and Away and equivalent checks for Over/Under and BTTS.
- Report proper scores. Calculate multiclass Brier score and log loss on the full probability vectors.
- Compare reference models. Use league base rates, a goals-only Poisson model and margin-free closing odds on the same fixtures.
- Inspect subgroups. Check league, season, favourites, underdogs, promoted teams and matches with unusual lineup uncertainty.
A model can predict average goals correctly while misallocating probability across exact scores. It can also fit exact scores while producing poorly calibrated 1X2 probabilities. Evaluate the outputs that will actually be used. For exact-score log loss, evaluate the probability of the actual observed score, including scores outside the displayed grid: loss = −log P(X = observed home goals, Y = observed away goals). A single “all other scores” category instead produces a coarsened categorical log loss. It can be reported separately, but it does not measure the probability assigned to each exact score in that tail.
See football probability calibration, Brier score and log loss for the full testing framework.
Limitations of converting xG into fair odds
The output is conditional on the projected xG
A precise probability calculation cannot repair biased λ inputs. Forecast error in xG is usually more important than rounding error in the Poisson formula.
Expected value does not determine variance
Poisson assumes variance equals the mean. Teams with the same projected xG can generate different score distributions if one creates many small chances and the other relies on a few high-value opportunities.
Home and away goals may be dependent
Score state, tactical reactions and shared match conditions can link the teams’ scoring processes. Dixon-Coles adjusts only four low-score cells; bivariate models use a broader dependence structure.
Scoring intensity changes during the match
A fixed λ averages over 90 minutes. Red cards, substitutions, fatigue and a team protecting a lead violate the constant-rate assumption.
Shot xG is conditional on shots occurring
Traditional xG evaluates observed attempts. It does not directly value possessions that fail to produce a shot or forecast the future number of shots. Pre-match modelling must account for both shot volume and shot quality.
Fair odds are not automatically profitable odds
A model may be internally coherent and still be less accurate than the market. Offered prices also include margin, movement, limits and execution constraints.
Retrospective fair odds answer a different question
Converting completed-match xG into probabilities can help assess whether the score was unusual relative to the chances. It cannot establish what should have been known before kickoff.
Use the correct label. “Fair odds from projected xG” is a forecast. “Outcome probabilities from realized shot xG” is a retrospective chance-quality analysis.
Practical checklist
- Confirm that the inputs are pre-match projected xG, not completed-match totals.
- State the xG provider, sample window and forecasting method.
- Use λhome and λaway in one coherent score matrix.
- Calculate derived markets by summing matrix cells.
- Use 1 ÷ probability for model-implied fair decimal odds.
- Estimate or validate Dixon-Coles ρ rather than choosing it to improve one example.
- Test the full pipeline chronologically and out of sample.
- Report calibration and proper scores, not only ROI or correct picks.
References
- Mead, J., O’Hare, A. & McMenemy, P. (2023). Expected Goals in Football: Improving Model Performance and Demonstrating Value . PLOS ONE, 18(4), e0282295.
- Cavus, M. & Biecek, P. (2022). Explainable Expected Goal Models for Performance Analysis in Football Analytics . arXiv:2206.07212.
- Wheatcroft, E. & Sienkiewicz, E. (2021). A Probabilistic Model for Predicting Shot Success in Football . arXiv:2101.02104.
- Davis, J. & Robberechts, P. (2024). Biases in Expected Goals Models Confound Finishing Ability . arXiv:2401.09940.
- 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.
- Gneiting, T. & Raftery, A. E. (2007). Strictly Proper Scoring Rules, Prediction, and Estimation . Journal of the American Statistical Association, 102(477), 359–378.