Betting With xG: Turning Data Into a Repeatable Method
A working method built on expected goals: which fields to collect, how to rate a fixture, converting two expected totals into a full grid of scorelines with Poisson, comparing the result with the coupon, backtesting it honestly, and the log fields that tell a bad estimate from bad luck.
Knowing what expected goals measures is not a method. A method is a fixed sequence that starts with published chance data and ends in a stake or a refusal, and gives the same answer whoever runs it on the same fixture. This page is that sequence: which fields to keep, how they become two numbers for a match, how those two numbers price every market, and how to tell afterwards whether a losing month was a bad estimate or bad luck. What the metric counts is covered in expected goals explained.
In short.
- Keep chance data at match level with penalties and set pieces split out, and never mix two providers.
- Shrink small samples towards the league baseline; ten matches is not ten matches' worth of evidence.
- Two expected totals plus a Poisson grid price every market on the coupon at once — here 55.2% home, 23.5% draw, 21.2% away.
- Against a board of 1.75, 3.90 and 5.00 only the away price qualifies, at +6.2%, worth ₦31 on a ₦500 stake.
- A 6% edge needs roughly 2,500 bets to clear two standard deviations, so the log has to diagnose the method before profit can.
What the method has to output
Five outputs, in order, each checkable on its own: a chance rating per club and venue; two expected totals for the fixture; a probability for every scoreline; a fair price per market; and a decision. If you cannot show those intermediate numbers a week later, it is not a method.
The general machinery is set out in building your own prediction model. What changes here is the input: chances created and allowed rather than goals scored and conceded, a quieter signal that needs different housekeeping.
Collecting the data: the fields worth keeping
One row per club per match. Every field below exists because leaving it out breaks a later step.
| Field | Why it is kept |
|---|---|
| Date, venue, opponent | Ratings are venue-specific and later weighted by opponent |
| xG created and allowed | The raw signal, before any adjustment |
| Penalty xG, both directions | Penalties are awarded, not manufactured; strip them before rating an attack |
| Set-piece xG, both directions | A side living on corners collapses when one centre-back is missing |
| Minutes level, ahead, behind | Chances built at 2-0 down describe an opponent that stopped defending |
| Red cards, with the minute | Everything after a dismissal belongs to a different match |
| Provider | Two models value the same shot differently; a mixed column is unusable |
| Goals scored and conceded | Not an input — the audit trail for the gap between chances and results |
The last row is the one people drop. Goals kept alongside chances measure whether a club has been over- or under-performing, and that gap is what the approach rests on: prices follow results, and results wander away from chances for months.
Turning a season of chances into two numbers
Work from non-penalty xG per match, at the venue in question, over the last ten matches, then pull the figure back towards the league baseline: 0.7 of your measured figure plus 0.3 of the baseline. That single step removes most of the damage a hot fortnight does. Take a division where home sides average 1.42 non-penalty xG a match and away sides 1.18.
| Input | Last 10, at venue | After shrinking | Against baseline |
|---|---|---|---|
| Home club, chances created at home | 2.05 | 1.86 | 1.86 ÷ 1.42 = 1.31 |
| Away club, chances allowed away | 1.29 | 1.33 | 1.33 ÷ 1.42 = 0.94 |
| Away club, chances created away | 1.01 | 1.06 | 1.06 ÷ 1.18 = 0.90 |
| Home club, chances allowed at home | 1.07 | 1.10 | 1.10 ÷ 1.18 = 0.93 |
Multiply the baseline by the two figures that meet in the fixture: the home side is expected to create 1.42 × 1.31 × 0.94 = 1.75, the visitors 1.18 × 0.90 × 0.93 = 0.99. As far as the method is concerned, that is the whole match.
The grid of scorelines
Goal counts follow a Poisson shape closely enough to price with: for an expected 1.75, the chance of exactly k goals is 1.75 to the power k, divided by k factorial, times e to the power −1.75.
| Goals | 0 | 1 | 2 | 3 | 4 |
|---|---|---|---|---|---|
| Home, 1.75 expected | 17.4% | 30.4% | 26.6% | 15.5% | 6.8% |
| Away, 0.99 expected | 37.2% | 36.8% | 18.2% | 6.0% | 1.5% |
Every scoreline is one row multiplied by one column: 1-0 is 0.304 × 0.372 = 11.3%, 2-1 is 0.266 × 0.368 = 9.8%. Add the cells a market settles on and you have its probability.
Reading the markets off the grid
Four markets from the same two numbers.
| Market | Your probability | Fair price | Coupon | Expected value |
|---|---|---|---|---|
| Home win | 55.2% | 1.81 | 1.75 | −3.4% |
| Draw | 23.5% | 4.25 | 3.90 | −8.2% |
| Away win | 21.2% | 4.71 | 5.00 | +6.2% |
| Over 2.5 goals | 51.6% | 1.94 | 1.90 | −2.0% |
Example. The three-way board of 1.75, 3.90 and 5.00 implies 57.1%, 25.6% and 20.0%, adding to 102.8%. Removing that mark-up proportionally, as in turning odds into probability, leaves the market on 55.6%, 25.0% and 19.5%. Your grid says 21.2% for the away side against the market's 19.5% — a gap of 1.7 points, which at 5.00 is 0.212 × 5.00 − 1 = +6.2%. On ₦500 that is ₦31 of expected value and a ₦2,500 return if it lands.
Notice how narrow that is. A gap under two percentage points is the whole bet, which is why the threshold matters more than the selection: below about 5% the edge is smaller than the error in your inputs. Most fixtures produce four passes. What the figure means across a season is in expected value explained.
Where plain Poisson is wrong
Four faults, none fatal, all worth knowing before you trust the output.
- Goals are not independent. A 1-0 changes how both sides play, so the raw grid slightly under-counts draws. Test a small uplift on the draw against your own history rather than assuming one.
- There is no game state. It prices ninety neutral minutes, not a match with a red card on the hour.
- Chance data is thin. Ten matches hold roughly 100 shots a side, and a few penalties move a rating by a tenth of a goal.
- Context sits outside the numbers. A third match in eight days, a settled title race or a missing goalkeeper appear nowhere in an xG table.
Important. Because the grid prices every market at once, it invites several bets on one fixture. They are the same opinion four times: if 1.75 is wrong, the result, the total and the scoreline selections are all wrong together, and one match takes four stakes with it.
The sample a fair test actually needs
Test on history under two rules: use only data that existed before each kick-off, rebuilding ratings match by match rather than from the finished season, and record the price you could actually have taken.
Then be honest about what the result proves. Say 300 bets at ₦500, average price 3.20, an edge the model puts at 6%. Expected profit is 0.06 × 300 = 18 units, or ₦9,000. One bet at those numbers has a standard deviation of about 1.51 units, so 300 carry 1.51 × √300 = 26.1 units, or ₦13,050.
A ₦9,000 expectation with a ₦13,050 spread means a good method finishes the test in the red often enough that the test decides nothing. Set expected profit equal to two standard deviations — 0.06n = 3.01√n — and n comes out near 2,500 bets. That is why closing-line comparisons, the subject of closing line value, reach a verdict years before the balance does.
The log that tells a bad estimate from bad luck
A standard bet log records what you staked. An xG log also records what you predicted, so the prediction can be judged separately from the result.
| Logged field | What it answers afterwards |
|---|---|
| Your two expected totals | Did the fixture behave as forecast? |
| The match's actual xG, both sides | Was the estimate close, whatever the score? |
| Your probability and fair price | Feeds the calibration check |
| Price taken, with the time | Separates a good bet from a good price |
| Closing price | The early skill signal |
| Team news after you staked | Explains a wrong estimate without excusing it |
The diagnosis runs on three comparisons. First, your predicted totals against the match's actual xG: within about 0.3 goals and the fixture behaved as forecast while the finishing did not, which is variance. Out by 0.8 or more in the same direction repeatedly and the estimate is wrong — usually an attack rated on set-piece or penalty chances it cannot repeat.
Second, the price. Taking 4.05 on a selection that closes at 3.85 is 4.05 ÷ 3.85 − 1 = +5.2% against the market, and a run of those alongside a losing balance is the best evidence you have that the sample is short rather than the method broken.
Third, calibration. Group settled bets by the probability you assigned: the band priced at 25–35% should land near 30% of the time. A shortfall inside one band is a model fault; one spread evenly across every band is noise. The fields every log needs are in how to track your bets.
Where to look. Chance figures, form and line-ups sit together on our match pages, which is where a fixture gets rated before the coupon is opened.
FAQ
How many matches of xG history do I need before rating a club?
Eight at the venue is the practical floor and ten is comfortable, provided the figures are shrunk towards the league baseline. Below eight, one match with three big chances moves the rating by more than the edge you are hunting, so the estimate measures a fortnight rather than a team.
Should the model use xG or actual goals as its input?
Chances, because goals are the noisier of the two: a side can create 2.1 and score none. The exception is a club whose finishing has beaten its chances over thirty-plus matches, where part of the gap may be skill, and a chances-only rating will keep underestimating it.
Why does my grid disagree with the coupon so often?
Usually because the coupon knows something a ten-match window does not: a suspension, a rested squad, a manager change. Treat a large gap as a question rather than an opportunity — check the team news first, and pass if the disagreement has no explanation you can name.
This article is for information only and is not an inducement to gamble. Betting involves the risk of losing money — never stake more than you can afford to lose. 18+. If gambling stops being entertainment, read our responsible gambling guide and seek help.