
A football forecast can give each selection a plausible-looking percentage while contradicting itself across the page. If the home side has a 52% chance of winning, its chance of winning or drawing cannot be 48% under the same match definition. Checking these relationships is a useful first step before comparing accuracy records or discussing prices.
Football prediction probability consistency means that estimates for related outcomes can fit together within one account of the match. The practical checks concern the information cutoff, the period being forecast and the events included. You can perform several without knowing how the forecasting model was built.
The exercise below uses invented numbers throughout. It audits the structure of a forecast and supplies questions to ask its publisher. It provides no estimate for a real fixture and makes no claim about any service’s historical performance.
Start by putting the forecasts on the same footing
Write down the fixture, publication time and period beside every percentage. A pre-match home-win forecast and an estimate issued after a sending-off describe different information sets. Comparing them as though they were simultaneous can create an apparent contradiction that disappears when the timestamps are restored.
The settlement period matters just as much. Winning after regulation time and qualifying after extra time or penalties are different events. A team can draw the match and still advance. Likewise, a first-half goals forecast belongs to a different period from a full-match goals forecast.
For this article, all examples concern the same fictional match, the same pre-match information and regulation time including stoppage time. Abandonment and operator-specific settlement conditions are outside the invented model. Before applying the examples to actual selections, read their definitions and establish whether those conditions change the comparison.
Check the complete match-result split
Home win, draw and away win form three separate result groups for a completed match under that definition. A forecast assigning probabilities to all three must account for the full set of possibilities. Stanford’s probability notes explain the total-mass and complement rules underlying this check. Stanford probability notes.
Suppose a page displays home win at 52%, draw at 27% and away win at 26%. The total is 105%. Those figures cannot be the exact probabilities of the three result groups within one distribution. Before deciding what failed, establish what the page is presenting.
Three independently copied bookmaker prices, converted to reciprocals, have a different interpretation from a model’s probabilities. A set of raw implied figures can include a pricing margin. A service may also show an undocumented confidence score that was never intended to be a probability. Ask for the meaning of the numbers before normalising them.
Dividing each displayed figure by 105 would force the total to 100%, but it would also invent a correction method. It would not show which estimate was wrong or recover the publisher’s intended forecast. Keep the original values and record the question they raise.
Use containment to test double chance and goal thresholds
Some relationships follow from which outcomes are included. Every home win is also a home-win-or-draw result. Every match with at least four goals also has at least three goals. Under compatible definitions, the smaller set cannot carry more probability than the set containing it.
For a fictional forecast of home win 52%, draw 27% and away win 21%, the home-or-draw probability is 79%. The away-or-draw probability is 48%. Adding all three double-chance probabilities would produce 200%, because those selections overlap. That sum is expected under these definitions and should not be mistaken for the match-result split.
Goal lines provide a second useful test. An estimate of 45% for over 3.5 goals and 40% for over 2.5 goals reverses their containment relationship. Check whether the lines use the same period, whether one value is conditional on a live score and whether the displayed labels match the underlying data.
These checks identify a specific incompatibility. They do not tell you whether either underlying estimate is close to the true match probability. A forecast can satisfy every relationship while remaining poorly informed.
Build a small score ledger to see the relationships
The following teaching distribution has only eight possible scores. Scores outside the table receive zero probability by construction. This is deliberately unrealistic as a forecasting model, but it lets you trace every calculation without an unreported tail of possible results.
| Home–away score | Assumed probability | Result group |
|---|---|---|
| 0–0 | 8% | Draw |
| 1–0 | 18% | Home |
| 0–1 | 12% | Away |
| 1–1 | 17% | Draw |
| 2–0 | 12% | Home |
| 0–2 | 8% | Away |
| 2–1 | 15% | Home |
| 1–2 | 10% | Away |
The entries sum to 100%. Home wins account for 45%, draws for 25% and away wins for 30%. Home or draw is therefore 70%. No choice about how to distribute probability between those eight scores remains once the table is fixed.
Over 1.5 goals collects 1–1, 2–0, 0–2, 2–1 and 1–2, giving 62%. Over 2.5 goals collects only 2–1 and 1–2, giving 25%. BTTS Yes collects 1–1, 2–1 and 1–2, giving 42%.
The point of this ledger is traceability. If a page displayed this exact score distribution alongside an over-2.5 probability of 55%, you could name the discrepancy precisely. Either the score table is incomplete, the markets come from different forecasts or the aggregation needs explanation.
BTTS and over 2.5 goals can legitimately disagree
The ledger gives BTTS Yes a higher probability than over 2.5 goals. That is compatible with its assumptions because the 17% allocated to 1–1 contributes to BTTS and stays below the three-goal threshold. In a different distribution, scores such as 3–0 and 0–3 could push over 2.5 above BTTS.
Neither selection generally contains the other. Treating their probabilities as interchangeable discards the distinction between goals scored by one team and goals shared between both. The existing BTTS analysis guide discusses that difference when reviewing team evidence.
There is also a sharper consistency check. For completed regulation-time scores, BTTS Yes and under 2.5 goals can occur together only at 1–1. In the teaching table, that intersection is exactly 17%. If a full distribution assigned 10% to 1–1 while separately claiming a 20% probability for that same intersection, the two claims would conflict.
You do not need to assume independent team scoring to make this observation. It follows from identifying the scores that satisfy both conditions. The assumption needed for multiplying two separate probabilities is a different question.
Check a combined forecast against its component events
Imagine a publisher gives home win 45%, BTTS Yes 42% and home win with BTTS Yes 50%. The combined claim cannot fit those component estimates. Both conditions must occur, so its probability cannot exceed either individual probability.
More generally, the overlap between two events is bounded above by the smaller probability and below by the larger of zero or their sum minus 100%. These bounds follow by accounting for the overlap in the addition rule. OpenStax presents that rule and distinguishes it from multiplication under independence. OpenStax probability rules.
For the 45% and 42% components, an overlap between zero and 42% is mathematically possible before more information is supplied. Our score ledger narrows it to 15%, because 2–1 is its only home-win-and-BTTS outcome.
Multiplying 45% by 42% would produce 18.9%. That figure imports an independence assumption which the ledger does not satisfy. The complete distribution supplies a direct answer, so substituting a multiplication shortcut would change the model.
Separate display rounding from a material conflict
A result split displayed as 34%, 33% and 32% might reflect values rounded separately from more precise underlying estimates. Ask for the unrounded numbers and the display policy. A difference caused by rounding should be reproducible from that policy.
For ordinary nearest-whole-percentage rounding, each displayed value differs from its underlying value by at most half a percentage point. Across three values, the absolute total discrepancy is therefore bounded by 1.5 percentage points. That is a conservative arithmetic bound, not permission to overlook any unexplained difference.
A five-point excess in a three-way split cannot be explained solely by that rounding rule. Nor can rounding plausibly explain a large reversal between nested goal thresholds. Preserve the exact displayed values, because repeated manual rounding can make a later review less informative.
Keep internal consistency separate from forecasting skill
A coherent set of probabilities could assign a favourite a 90% win chance for weak reasons. Its calculations might agree perfectly while its estimates perform badly across comparable fixtures. Assessing that problem requires a dated forecast archive, outcomes and a suitable evaluation method.
Equally, an inconsistent display does not establish fraud or explain how the underlying model performs. A stale widget, a label error or two different update times could be responsible. Those possibilities need evidence. Record what you observed and ask which version of the forecast the service intends readers to evaluate.
Price comparisons introduce further requirements. Raw decimal odds reciprocals are useful pricing quantities, but comparing them with a model requires a stated treatment of margin and relevant costs. Passing a consistency check supplies none of that information. It also provides no reason to increase a stake.
A review note that another reader can reproduce
Save the page URL, observation time, fixture identifier, forecast cutoff and period. Copy the original percentages and their labels. State the relationship being tested, show the arithmetic and distinguish missing information from a demonstrated contradiction.
A useful note might say that a same-time pre-match home-win estimate exceeds the home-or-draw estimate by four points, with screenshots attached and the publisher’s explanation pending. That is more actionable than calling an entire prediction site unreliable after one unexplained display.
If an explanation arrives, retain both versions. Record whether the cause was rounding, incompatible periods, different forecast timestamps or a genuine aggregation error. A corrected page should remain distinguishable from the original evidence.
When comparing services, apply the same procedure to each. A consistent record of checks makes it possible to separate a corrected presentation fault from a recurring problem. Once the definitions and arithmetic agree, you can move on to whether the forecasts actually help readers understand matches.
评论 0
暂无评论。