How Accurate Are Football Predictions? A Data-Driven Study
How accurate are football predictions, and what should an AI model achieve before its forecasts can be called reliable? Footiqo analysed 1,751 recorded matches from the Premier League, LaLiga, Bundesliga, Serie A and Ligue 1 in 2025/26. We compared the results with the most likely outcomes indicated by historical opening and closing betting odds, then examined probability calibration and Both Teams To Score (BTTS) predictions.
This is a benchmark for evaluating AI predictions, not a test of an AI model. We use historical bookmaker odds as a measurable pre-match reference. The findings show why a headline accuracy figure is not enough to judge a forecasting system.
Key findings
- Choosing the lowest closing 1X2 price beat always choosing a home win by 9.09 percentage points on the same 1,727 fixtures (53.21% versus 44.12%).
- The Bundesliga had the highest closing-odds 1X2 accuracy (57.19%); the Premier League had the lowest (48.53%).
- Opening and closing odds were virtually tied on 1X2 hit rate overall (53.27% versus 53.21%). This does not establish that opening prices are better probabilistic forecasts.
- Draws accounted for 25.42% of 1X2 sample matches, but the shortest closing price pointed to a draw in only five matches (0.29%).
- In the BTTS sample, the closing-odds selection beat the simple “always Yes” rule by 2.64 percentage points.
Why prediction accuracy needs a baseline
“Did the forecast identify the winner?” sounds like a simple test, but the answer depends on what was predicted. A home/draw/away (1X2) selection, a BTTS selection and an expected-goals estimate concern different events. They cannot be scored as though they were the same prediction.
Hit rate is the number of correct selections divided by all eligible selections. A model that gets 60 out of 100 calls right has a 60% hit rate, but that figure is of limited use if a simpler rule would have selected 62 correctly on those same matches. A fair comparison requires an identical fixture list, defined selection rules and a frozen pre-match information cut-off.
In our research, the benchmark is straightforward: select the outcome with the lowest available pre-match 1X2 odds. We compare that with a baseline that always predicts a home win. Football prediction examples on Scoredium provide a separate, practical context for seeing the kinds of pre-match calls that a prediction audit needs to preserve.
1X2 football prediction accuracy by league
Of the 1,751 matches in our five-league dataset, 1,727 had a unique shortest opening price and a unique shortest closing price. On this identical sample, the opening-odds rule selected 920 winners (53.27%) and the closing-odds rule 919 (53.21%). Simply selecting the home team would have been correct in 44.12%.
| League | Matches | Opening 1X2 | Closing 1X2 | Always home |
|---|---|---|---|---|
| Premier League | 375 | 49.07% | 48.53% | 42.40% |
| LaLiga | 375 | 53.87% | 53.60% | 49.07% |
| Bundesliga | 299 | 55.52% | 57.19% | 44.15% |
| Serie A | 375 | 54.40% | 54.13% | 39.20% |
| Ligue 1 | 303 | 54.13% | 53.47% | 46.20% |
| All five leagues | 1727 | 53.27% | 53.21% | 44.12% |
Source: Footiqo historical match results and 1xBet opening/closing 1X2 odds, 2025/26. All methods use the same fixtures. We excluded 24 matches where the minimum price was tied at opening or closing; no arbitrary tie-break was imposed.
Which league was easiest to predict?
The Bundesliga led this particular test at 57.19% with closing prices, compared with 48.53% in the Premier League, a difference of 8.66 percentage points. This describes how often the market’s single most likely result occurred; it does not mean that betting on Bundesliga favourites was profitable, nor does it measure an AI model’s success rate.
Did closing odds beat opening odds?
Not on the simple hit-rate measure: closing prices selected one fewer correct result across 1,727 games. The difference was just 0.06 percentage points. Later odds may still have better probabilities, even when the most likely outcome does not change. Assessing that question properly requires measures such as log loss or Brier score for all three outcomes, not hit rate alone.
The draw problem: 439 actual draws, only five draw selections
Within the analysed 1X2 sample, 439 matches ended in draws. Yet the draw had the lowest closing odds in just five. A rule that always chooses whichever of home, draw or away has the highest implied probability therefore almost never selects a draw.
That is not proof that the odds assigned draws too little probability. A 27% chance of a draw can be perfectly sensible even when a home win is more likely at 45%. It shows instead why a single picked outcome and the full probability distribution are different ways of judging a forecast.
Probability calibration: were the favourites as likely as the odds suggested?
Accuracy asks whether the selected result happened. Calibration asks whether outcomes assigned similar probabilities occurred with similar frequencies. If a system assigns 70% to a particular event across many matches, that event should happen in roughly 70% of them if the probabilities are well calibrated.
We converted the three closing 1X2 prices into implied probabilities, then normalised them to sum to 100% to remove the bookmaker overround. For each match we retained the probability of its most likely outcome and grouped these probabilities into ten-percentage-point bands.
Calibration gap: observed hit rate minus the mean normalised predicted probability, in percentage points.
| Probability band | Matches | Mean probability | Actual hit rate | Gap |
|---|---|---|---|---|
| 30-39% | 279 | 37.64% | 37.63% | -0.01 pp |
| 40-49% | 591 | 44.90% | 46.19% | +1.29 pp |
| 50-59% | 424 | 54.65% | 53.30% | -1.34 pp |
| 60-69% | 266 | 64.46% | 67.67% | +3.21 pp |
| 70-79% | 134 | 74.62% | 78.36% | +3.74 pp |
| 80-89% | 32 | 84.31% | 90.63% | +6.32 pp |
| 90-99% | 1 | 90.77% | 100.00% | +9.23 pp |
The 90–99% band contains only one match; its 100% observed rate is not a meaningful stand-alone estimate. The 80–89% band contains 32 matches, so that result is also particularly uncertain.
Across all 1,727 selected matches, the mean estimated probability was approximately 52.20% and the observed hit rate was 53.21%. The 30–39% band was almost identical (37.64% predicted, 37.63% realised, 279 matches). For selections in the 70–79% range, the observed rate was 78.36% against 74.62% estimated, although the sample was smaller (134 matches).
These figures suggest that predicted and observed frequencies were reasonably close in several bands. They are not a full calibration assessment of the three-way market, because the chart covers only the most likely outcome, and differences within a season can arise from ordinary variation.
Football’s expected goals (xG) statistics add useful context when explaining surprising results after a match. But post-match xG must not be treated as a pre-match input that an AI model could already have known.
How accurate are Both Teams To Score predictions?
BTTS is a binary question: will both teams score at least once? A correct 1X2 selection does not automatically count as a correct BTTS selection. This market therefore deserves its own accuracy test and its own baseline.
We selected “Yes” when its BTTS odds were shorter than “No”, and vice versa. Among 1,554 comparable matches, opening BTTS odds yielded 891 correct predictions (57.34%), while closing odds yielded 887 (57.08%). Always choosing BTTS Yes would have achieved 54.44% on the same fixtures.
| League | Matches | Opening BTTS | Closing BTTS | Always Yes |
|---|---|---|---|---|
| Premier League | 353 | 60.34% | 58.92% | 56.94% |
| LaLiga | 321 | 57.63% | 58.88% | 56.70% |
| Bundesliga | 293 | 59.73% | 60.41% | 61.77% |
| Serie A | 325 | 54.77% | 51.38% | 44.62% |
| Ligue 1 | 262 | 53.44% | 55.73% | 52.29% |
| All five leagues | 1554 | 57.34% | 57.08% | 54.44% |
Source: Footiqo and 1xBet opening/closing BTTS odds, 2025/26. We excluded 197 of 1,751 matches where Yes and No were tied for the shortest price at opening or closing. This 1,554-match BTTS sample is different from the 1,727-match 1X2 sample; the hit rates should not be compared as if they measured the same event or identical fixtures.
Why league-specific baselines matter
The “always Yes” rule would have hit 61.77% of Bundesliga games in the BTTS sample, slightly more than the 60.41% achieved by the closing-price selection rule. In Serie A, the opposite baseline mattered: always selecting BTTS No would have achieved 55.38%, compared with 51.38% for the closing-price selection. Neither observation establishes a durable betting edge, especially after accounting for price and any stake strategy.
You can explore league and team-level scoring patterns further through Footiqo’s free full-time BTTS statistics, including splits by team, home and away performance.
What would an AI football prediction model need to beat?
Our analysis provides a real-world reference, rather than a ranking of AI tools. A useful AI model should be tested against the corresponding market benchmark using exactly the same fixtures and forecast deadlines. It should disclose whether it predicts only an outcome or supplies probabilities for every outcome.
- Freeze every forecast before kick-off: save the original prediction, probability, model version and timestamp, without silently overwriting earlier calls.
- Use the same fixture list: distinguish all-fixture coverage from selective predictions, and count abstentions explicitly.
- Publish separate markets: 1X2, BTTS, totals and xG-based estimates cannot be merged into a single unexplained accuracy number.
- Compare like with like: use both a simple baseline and the corresponding pre-match odds, and disclose any exclusions.
- Report probability quality: add calibration, a proper scoring rule and sample counts alongside the categorical hit rate.
- Keep mistakes visible: publish the complete audit trail and describe any later corrections.
A public record is useful because readers can check what was predicted rather than rely on selected winning examples. Scoredium’s prediction leaderboard is one example of a publicly visible prediction-ranking format; any model-by-model comparison still needs to establish whether its underlying picks, markets and grading rules are comparable with the benchmark used here.
A season-long scorecard may offer evidence about specific fixtures and prediction types, but it cannot establish that any model is permanently the best. Teams, line-ups, competitions and model versions change. AI-versus-human comparisons likewise require consistent information deadlines and grading criteria.
Methodology, sample and limitations
- Source: Footiqo historical match database; observed full-time match results and recorded 1xBet opening and closing pre-match odds.
- Season: 2025/26 (the five major European domestic leagues). The available match counts were 380 Premier League, 380 LaLiga, 306 Bundesliga, 380 Serie A and 305 Ligue 1, giving 1,751 records with results and market coverage.
- 1X2 inclusion: complete full-time results and positive, valid opening/closing prices for H/D/A. We required a unique lowest price in both opening and closing markets; 24 price ties were excluded, leaving 1,727 matches.
- BTTS inclusion: complete results and valid opening/closing Yes/No prices; 197 ties excluded, leaving 1,554 matches. BTTS Yes means both teams scored at least once in full time.
- Hit-rate rule: the lowest decimal odds represent the selected outcome; each correct result counts once. “Always home”, “Always Yes” and “Always No” baselines use the same eligible matches as their market comparisons.
- Calibration: normalised closing 1X2 probabilities calculated from all three decimal prices, then assessed only for the shortest-priced outcome. Ten-percentage-point probability bands were used without post-hoc adjustment.
- Coverage limitation: the extracted Ligue 1 dataset contains 305 recorded matches, so this research reports the observed sample rather than asserting that every fixture in that league is represented.
- Odds limitation: opening and closing represent the first and final recorded pre-match prices in Footiqo’s data. This study does not independently validate an exact timestamp for each bookmaker quote.
- Interpretation: a higher hit rate is not proof of profitability, positive expected value or statistically significant superiority. Exclusion of tied prices may affect representativeness, particularly for BTTS.
For researchers wishing to reproduce or extend this work, Footiqo provides free football dataset samples and a broader historical football database with match results and recorded betting odds.
Conclusion: prediction accuracy is more than a percentage
For 2025/26, the closing-odds 1X2 benchmark selected the correct result in 53.21% of 1,727 comparable Big Five fixtures, while closing BTTS selections were right in 57.08% of a separate 1,554-match sample. Those figures are meaningful references, but they are not proof that any AI model reaches, exceeds or falls short of them.
The bigger lesson is methodological: declare what is predicted, use the same matches, compare with a baseline, examine probabilities and keep failures visible. Platforms such as Scoredium make public football prediction records part of the wider conversation; Footiqo’s historical data provides an independent benchmark against which rigorously archived forecasts can be assessed.
Frequently asked questions
What is a good accuracy rate for AI football predictions?
It depends on the target market, fixture mix and baseline. In our 2025/26 sample, the closing-odds most-likely 1X2 outcome was correct 53.21% of the time. An AI model would need to be evaluated on the same fixtures and criteria before any comparison was meaningful.
Is a 60% football prediction hit rate good?
Possibly, but not by itself. A 60% rate on a sample dominated by heavy favourites might be less impressive than it appears. Always check the number of matches, the type of prediction, its difficulty and the corresponding baseline.
Why do bookmakers rarely predict draws as the most likely outcome?
A draw can have a substantial probability without being the single most likely result. In our 1X2 sample, draws occurred 439 times, although the draw was the shortest-priced closing outcome in only five fixtures. This alone does not imply that draw probabilities were mispriced.
Are closing odds more accurate than opening odds?
Not necessarily by categorical hit rate. In this 2025/26 sample, opening 1X2 odds selected 920 correct outcomes and closing odds 919. A proper probability-quality comparison would also assess all outcome probabilities using measures such as log loss or Brier score.
Does a higher prediction hit rate mean profitable betting?
No. Profitability also depends on the odds obtained, bookmaker margin, stake sizing and whether the prices were available at bet placement. This study measures outcome-selection accuracy, not betting returns.
Using this research
Publishers, journalists and researchers are welcome to cite the statistics and charts in this study. Please credit Footiqo and link to this page as the original source of the 2025/26 analysis.
Research: Footiqo.com. Statistics calculated from historical match results and recorded pre-match odds. Values are rounded to two decimal places; minor differences may appear when rounding.
Contact Us
If you have any questions, or would like us to add specific leagues or betting markets, please contact us via our contact form or email us at support@footiqo.com. We’ll get back to you as soon as possible.
Enjoying these stats?
If you find this useful, share it with your mates on social media.