How Accurate Were the 2026 World Cup Betting Odds?

This completed study compares probabilities implied by pre-match betting odds with what happened across all 104 matches of the 2026 World Cup. It does not count betting wins or promote correct-score picks. It asks a narrower statistical question: did outcomes occur about as often as the probability model said they should?

Data snapshot: 20 July 2026, after the final. All 104 matches are settled, and every figure on this page is frozen at the complete-tournament state.

Primary 1X2 and Over/Under inputs are de-vigged pre-match prices from one tracked bookmaker (Duel); correct-score comparisons use Oddschecker where captured. This is not a multi-bookmaker market-consensus study.

Final conclusion
  • The final score appeared inside the model’s top five in 56 of 104 matches, versus 55.7 expected.
  • The model’s highest-probability exact score occurred 18 times, versus 14.3 expected.
  • The average probability assigned to the actual score was 8.1%, close to the 7.9% collision benchmark.
  • Exact Poisson-binomial tests place every top-k hit count inside its 95% prediction interval.
  • 1X2 Brier score and log loss beat a uniform baseline by roughly 25% and 23%; Over/Under skill is small.
  • The remaining gaps are consistent with ordinary tournament variance.

World Cup 2026 odds accuracy: final results

104

settled matches

17.3%

actual score ranked #1

53.8%

actual score inside top five

8.1%

mean probability of actual score

49.2%

mean probability of actual 1X2 outcome

51.9%

mean probability of actual O/U side

Metric Observed Expected benchmark Difference Current reading
Top-1 exact-score coverage 17.3% 13.7% +3.6pp Above expectation; exact test z = +1.06, p ≈ 0.18
Top-3 exact-score coverage 39.4% 36.4% +3.0pp Broadly close (z = +0.65)
Top-5 exact-score coverage 53.8% 53.5% +0.3pp Essentially at expectation (z = +0.06)
Probability assigned to actual score 8.1% 7.9% +0.2pp Very close
Probability assigned to actual 1X2 outcome 49.2% 48.0% +1.2pp Close in aggregate
Probability assigned to actual O/U 2.5 side 51.9% 51.8% +0.1pp Near-exact; identical on the direct-only subset

“Observed” and “expected” do not mean the same thing for every row. Top-k rows compare the number of matches in which the final score landed inside a ranked group with the total probability assigned to that group. The other rows compare the mean probability assigned to the realized outcome with a collision-probability benchmark.

What do these World Cup probability metrics mean?

Top-1 coverage How often the final score was the single highest-probability scoreline before kickoff.
Top-3 coverage How often the final score appeared among the three highest-probability scorelines.
Top-5 coverage How often the final score appeared among the five highest-probability scorelines.
Actual score probability The average pre-match probability assigned to the exact score that occurred.
Actual 1X2 probability The average probability assigned to the realized home/draw/away result.
Actual O/U probability The average probability assigned to the realized Over or Under 2.5 side.

49.2% is not a 49.2% 1X2 hit rate. It is the average probability the tracked odds assigned to whichever 1X2 outcome occurred. The odds could assign 70% to one winning favourite and 20% to one winning underdog; their average actual-outcome probability would be 45%.

Likewise, an 8.1% actual-score probability is not the percentage of correct-score predictions. The model’s single top score was correct in 17.3% of matches (18 of 104). The 8.1% figure measures how much probability was assigned to the realized score across all matches, including scores ranked second, tenth or lower.

How the World Cup tracker converts odds into score probabilities

Each match begins with a pre-match snapshot of the 1X2 and Over/Under market. The tracker then applies the same sequence before the result is known.

  1. Capture the odds. Record home, draw, away, Over 2.5 and Under 2.5 prices from the tracked bookmaker (Duel) before kickoff.
  2. Remove the bookmaker margin. Convert the prices to raw implied probabilities and normalize each market separately.
  3. Infer goal rates. Find λhome and λaway values that reproduce the fair 1X2 and total probabilities as closely as possible.
  4. Build a score matrix. Use Poisson or Dixon-Coles to assign probability to 0-0, 1-0, 1-1, 2-1 and every other score.
  5. Freeze the pre-match distribution. Store the ranked scorelines and calibration benchmarks.
  6. Overlay the result. After the match, record the regulation-time score, its probability and its rank.

The method is a translation of betting prices, not an independent team forecast. 1X2 and totals are inputs. The exact-score matrix is the model layer that converts those inputs into a coherent distribution.

The primary 1X2 and totals inputs in this study come from Duel. They should be interpreted as one bookmaker’s prices — typically close to, but not identical with, a multi-book market average, a sharp closing line or an exchange midpoint.

The full process is documented in the ImpliedScore methodology and the guide to inferring goal rates from betting odds.

View all 104 match cards, including pre-match odds, λ values, top scorelines, final scores and per-match calibration fields.

Open the live tracker

Did the most likely World Cup scores occur often enough?

Before each match, the score matrix assigns a probability to every exact result. If the highest-ranked score has a 14% probability, it should occur in roughly 14% of comparable matches—not most of the time.

Rankings use unrounded probabilities, and Top-1, Top-3 and Top-5 always mean exactly one, three and five cells: exact ties are broken by a predeclared stable rule applied identically to the observed and expected sides. The full boundary policy is documented in the most-likely-score analysis.

Top-1: 18 observed versus 14.3 expected

The model’s first-ranked score occurred in 18 of 104 matches. The per-match top-score probabilities imply 14.29 expected occurrences. The observed excess is 3.7 matches.

That is directionally positive for the model, but not decisive. The exact Poisson-binomial test — each match keeps its own top-score probability — gives a standard deviation of 3.51 hits, so the excess is z = +1.06, with a one-sided p-value of 0.18 for observing 18 or more hits. The model’s own 95% prediction interval is 8 to 21 hits, and 18 sits inside it.

A naive binomial approximation with one shared hit probability overstates this excess — on the 100-match interim sample it suggested roughly z ≈ 1.25 — because it understates the expected count when per-match probabilities vary. The exact test is the one to trust.

Top-3: 41 observed versus 37.8 expected

The actual score appeared inside the three highest-probability outcomes in 41 matches, against 37.82 expected from the per-match probabilities (z = +0.65). The 95% prediction interval is 28 to 48 hits.

The one-sided Poisson-binomial p-value for 41 or more hits is 0.29 — comfortably inside ordinary variation.

Top-5: 56 observed versus 55.7 expected

Top-five coverage is the closest result: 56 hits against 55.69 expected (z = +0.06). The 95% prediction interval is 46 to 66 hits; the observed count sits essentially at the centre of the model’s own distribution.

The correct target is not 100%. A top-five group containing 53.5% total probability should miss about 46.5% of the time. A high miss count is not evidence of model failure when the model itself assigned substantial probability outside the top five.

Why the expected actual-score probability is a collision probability

Suppose one match has exact-score probabilities p1, p2, … across the complete matrix. If the model is calibrated, score s occurs with probability ps. When it occurs, the probability recorded for the actual score is also ps.

The expected probability assigned to the realized score is therefore:

E[pactual score] = Σs ps × ps = Σs ps²

This is called a collision probability: the chance that two independent draws from the same distribution produce the same category.

Across all 104 World Cup matches:

Observed mean pactual score = 8.1%    vs    Expected mean Σps² = 7.9%

The 0.2-point difference is small. It suggests that the realized scorelines have, in aggregate, not been systematically concentrated in much more or much less probable matrix cells than the model expected.

This is not a full calibration test. These averages are low-dimensional aggregate self-consistency diagnostics: the match condition Σp(r − p) = 0 can hold even when the distribution is wrong, because errors in some scorelines offset errors in others. Exact-score log loss, Brier-type scores and grouped reliability analysis are still needed.

Were the World Cup 1X2 probabilities accurate?

For a match with fair probabilities H, D and A, the expected probability assigned to the outcome that occurs is:

H² + D² + A²

The tracker reports:

Observed actual-outcome probability = 49.2%    vs    Expected collision benchmark = 48.0%

The aggregate difference is +1.2 percentage points. This is consistent with broadly reasonable 1X2 probabilities, but it does not reveal whether home wins, draws and away wins are each calibrated separately.

What a stronger 1X2 analysis still requires

  • separate reliability diagrams for Home, Draw and Away with broad pre-specified bins;
  • comparison with a pre-specified historical or pre-tournament base-rate forecast;
  • uncertainty intervals for the difference from a reference model;
  • published per-card odds timestamps alongside the stated margin-removal method.

Multiclass Brier score and log loss are now computed for the completed tournament and reported in Brier score and log loss below.

With only 104 matches in the completed tournament, fine 10-bin calibration curves would be unstable. Broad pre-specified bins or classwise regression should be used cautiously.

The evaluation framework is explained in football probability calibration, Brier score and log loss.

Were the Over/Under 2.5 probabilities accurate?

For a binary Over/Under market with fair Over probability O and Under probability U = 1 − O, the collision benchmark is:

O² + U²

The tracker reports a near-exact aggregate match:

Observed actual-side probability = 51.9%    vs    Expected benchmark = 51.8%

This is the closest aggregate result in the tracker. It does not mean the Over side itself has been perfectly calibrated. The same average can arise if one probability range is too high and another is too low.

There is a data-definition caveat: three match cards — France-Senegal, Argentina-Algeria and Ghana-Panama — use a team-total anchor because a quoted Over/Under 2.5 price was unavailable. On those cards, the O/U benchmark is model-derived rather than directly market-derived. Excluding them, the direct-only subset (n = 101) shows 51.9% observed versus 51.8% expected — identical to the combined figure at displayed precision. The headline agreement is not an artefact of mixing the two data types.

Brier score and log loss for the completed tournament

Proper scoring rules evaluate the full probability vector of every match, not only aggregate averages. The values below are computed from the frozen per-card de-vigged probabilities and the 104 regulation-time results. A uniform forecast — every outcome equally likely — serves as the trivially pre-specified baseline.

Market Score Uniform baseline Skill versus uniform
1X2 multiclass Brier score 0.497 0.667 +25.5%
1X2 multiclass log loss 0.843 1.099 +23.3%
Over/Under 2.5 Brier score 0.240 0.250 +4.0%
Over/Under 2.5 log loss 0.672 0.693 +3.0%
Exact-score log loss 2.799 — —

The 1X2 probabilities carried substantial information: roughly a quarter better than uniform by Brier score. The Over/Under margin over a coin flip is small, and honestly so — the de-vigged totals probabilities clustered close to 50%, so little separation was available to reward.

Exact-score log loss is tail-sensitive. The 4-6 third-place match alone, with a published probability of 0.02%, contributes −ln(0.0002) ≈ 8.5 to the sum; excluding that single card lowers the mean from 2.799 to 2.744. Both figures are reported deliberately: a proper score must charge the model for its tails, and one ten-goal match in 104 is what those tails look like when they materialize.

Scores are computed from per-card probabilities published at one- or two-decimal precision; recomputation from unrounded matrices could shift the third decimal place. A historical pre-tournament base-rate baseline remains future work.

What individual matches show—and why anecdotes are insufficient

Mexico 2-0 South Africa: a high-probability result

The tracker assigned 15.0% to 2-0, making it the second-highest scoreline. The direct correct-score market, where captured and de-vigged, assigned 13.1%. This is an example of a result landing in a dense, high-probability region of the matrix.

Sweden 5-1 Tunisia: a low-probability tail result

The final score had only about 0.5% model probability and ranked #21. Such a result is unusual, but not impossible. Across 104 matches, some low-probability scorelines should occur.

Norway 1-1 England: the top score occurred

The regulation-time score 1-1 was ranked first with 12.2% probability. Calling this a “successful prediction” without the 12.2% qualifier would be misleading: the same top score was expected to fail almost seven times out of eight.

France 4-6 England: the deepest tail of the tournament

The third-place match produced ten goals and finished ranked #51 in the pre-match matrix, with a published probability of 0.02%. One occurrence of such a scoreline in 104 matches is compatible with the tails the matrices assign — and it is exactly the kind of result that dominates tail-sensitive metrics such as exact-score log loss.

One spectacular hit or miss does not validate or refute a probability model. The relevant question is whether many outcomes, including rare ones, occur at rates consistent with the probabilities assigned before kickoff.

So, were the 2026 World Cup betting odds accurate?

The defensible current answer is: the de-vigged pre-match probabilities from the tracked bookmaker, and the score matrices derived from them, show no large aggregate inconsistency with the completed 104-match tournament. The aggregate metrics are low-dimensional self-consistency checks — now supplemented by exact Poisson-binomial tests and proper scoring rules — and they describe one bookmaker’s prices rather than the entire betting market.

Two different things are being evaluated. The 1X2 and Over/Under rows test whether the tracked bookmaker’s de-vigged input probabilities were calibrated; the exact-score rows test whether the Poisson/Dixon-Coles translation of those inputs into a score matrix was probabilistically coherent. Neither row type validates the other.

The strongest evidence is the closeness of:

  • top-five coverage: 53.8% observed versus 53.5% expected;
  • actual-score probability: 8.1% versus a 7.9% collision benchmark;
  • actual 1X2 probability: 49.2% versus 48.0%;
  • actual O/U probability: 51.9% versus 51.8%, identical on the direct-only subset.

Top-1 coverage ended above expectation—17.3% versus 13.7%—but the exact Poisson-binomial test places 18 hits at z = +1.06 with a one-sided p-value of 0.18 on the frozen final dataset: inside ordinary tournament variance.

These results also do not show that bookmakers predicted exact scores directly. Most exact-score probabilities in the tracker are generated by fitting a Poisson or Dixon-Coles matrix to margin-free 1X2 and totals odds. They test whether that translation is probabilistically coherent with the observed scores.

“Accurate” does not mean “profitable to bet against or with.” A well-calibrated market can still produce negative expected returns for bettors because offered odds include margin. Probability quality and betting profitability are separate questions.

What this World Cup analysis does not prove

It is one tournament

One hundred and four matches are enough for descriptive analysis but too few for precise calibration in narrow probability ranges or rare exact scores.

The fixtures are not identically distributed

Group matches, knockout matches, heavy favourites and balanced fixtures have different probability structures. Treating all matches as interchangeable hides that heterogeneity.

Knockout advancement is not the 1X2 outcome

Standard 1X2 odds settle on the score after regulation time unless stated otherwise. Extra time and penalty shootouts determine who advances, but they should not be substituted for the 90-minute score when evaluating a regulation-time market.

The exact-score model is constrained by the input markets

1X2 and Over/Under prices are used to fit λ. Good reproduction of those inputs is partly by construction. The independent empirical question is whether the resulting score matrix is calibrated against final scores.

Poisson and Dixon-Coles are approximations

Constant scoring rates, conditional independence and a four-cell dependence correction do not capture every tactical and match-state effect.

Correct-score market data are incomplete

Direct correct-score prices are shown only for matches where they were captured. A valid de-vig also requires a complete, mutually exclusive outcome set: captured score lists must include the “any other score” residual buckets, and partial lists must not be proportionally normalized as if they covered the full outcome space. No full tournament claim about the direct correct-score market should be made until coverage, completeness and timing are documented.

Odds timing matters

Opening, current and closing prices reflect different information. The same timing policy should be used across matches, or timing variation should be disclosed. Per-card capture timestamps are not yet published; until they are, the pre-match status of each snapshot rests on the documented capture protocol rather than on independently verifiable timestamps.

The margin-removal method matters

The tracker uses a transparent de-vigging procedure. Alternative methods can produce slightly different fair probabilities, especially for strong favourites and high-margin markets.

Aggregate agreement can hide classwise bias

Home probabilities may be too high while away probabilities are too low, yet average collision metrics can remain close. Outcome-specific diagnostics are needed.

What this final report adds — and what remains open

With the final played on 19 July 2026, the live page has been frozen into a complete tournament study. Relative to the live version, this report now includes:

  • exact Poisson-binomial tests and 95% prediction intervals for Top-1, Top-3 and Top-5 coverage;
  • multiclass Brier score and log loss for 1X2, binary scores for Over/Under 2.5 and exact-score log loss with an explicit tail treatment;
  • a direct-only Over/Under result with the three team-total fallback cards identified and excluded;
  • the frozen 104-match dataset table on the tracker page.

Open items that would strengthen the study further:

  • classwise Home/Draw/Away calibration with broad pre-specified bins and uncertainty intervals;
  • a descriptive group-stage versus knockout split (the knockout sample is small);
  • a paired Poisson versus Dixon-Coles comparison on identical pre-match inputs;
  • a direct correct-score subset with documented market completeness and timing;
  • bootstrap uncertainty intervals around the headline aggregates;
  • published per-card odds-capture timestamps;
  • a documented rule for how the tracker selects Poisson, Dixon-Coles and ρ.

A tournament of 104 matches is still too small to identify subtle biases reliably. Its main value is transparent, pre-match documentation and a worked demonstration of probabilistic evaluation—not a definitive verdict on all football betting markets.

Data freeze protocol

The study is complete. The rules below governed the live phase and now define the frozen state; they are retained so the final numbers can be audited against the stated process.

  • Every match was captured before kickoff and settled on the regulation-time score.
  • Odds source, model choice, ρ and λ values are stored on each card.
  • Pending matches were excluded from every aggregate until settled.
  • Model settings were not altered retroactively for completed matches.
  • The 104-match state of this page and the tracker is the frozen final version.
  • Any post-final correction will be disclosed in a changelog rather than silently applied.

References and data sources

  1. ImpliedScore. World Cup 2026 Market-Implied Score Tracker . Pre-match odds snapshots, score matrices, final-score overlays and aggregate metrics.
  2. FIFA. FIFA World Cup 2026 match schedule, fixtures and results . Official 104-match tournament schedule, 11 June–19 July 2026.
  3. ImpliedScore. Methodology: From Odds to a Score Probability Matrix.
  4. ImpliedScore. Football Probability Calibration: Brier Score, Log Loss and Reliability .
  5. Brier, G. W. (1950). Verification of Forecasts Expressed in Terms of Probability . Monthly Weather Review, 78(1), 1–3.
  6. Gneiting, T. & Raftery, A. E. (2007). Strictly Proper Scoring Rules, Prediction, and Estimation . Journal of the American Statistical Association, 102(477), 359–378.

Educational analysis only. This page evaluates probabilities and does not provide betting recommendations. Odds do not guarantee outcomes or profit. Betting involves financial risk and is intended only for adults in jurisdictions where it is permitted.

Scroll to Top