How our AI works, and how we test it against the bookmakers
Every pick on Daily Sport Pick comes from an automated system: three prediction models score each match and fixed rules decide what gets published. No tipster picks by hand. Below are the models, the rules, and the tests we run on both, including the ones we fail.
Our model against the bookmaker market
In September we compared our model’s probabilities with the probabilities implied by bookmaker odds, margin removed, on 3,519 matches between April and September 2026. Every prediction was made before kick-off. We used log-loss, a standard score for probability forecasts: lower is better, and confident mistakes are punished hard. As a reference we also scored a forecast that always predicts the average result.
| Match result, by competition group | Matches | Our model | Bookmaker odds | Always the average |
|---|---|---|---|---|
| Top 5 leagues | 553 | 1.025 | 1.004 | 1.081 |
| Eredivisie, Primeira Liga, Belgian Pro League, Championship, Brasileirão, MLS | 709 | 1.047 | 1.003 | 1.081 |
| European cup competitions | 468 | 1.099 | 0.926 | 1.051 |
| Second divisions and mid-size European leagues | 551 | 1.127 | 1.024 | 1.080 |
| All other competitions | 1,238 | 1.086 | 1.023 | 1.073 |
| All matches | 3,519 | 1.076 | 1.003 | 1.075 |
Source: Daily Sport Pick live predictions against bookmaker odds with the margin removed, April to September 2026.
- The bookmaker market was the better forecast in every market we tested (match result, over/under 1.5, 2.5 and 3.5 goals, both teams to score) and in every competition group.
- In the top leagues our model came close to the market. In European cups and smaller leagues it scored worse than always predicting the average.
- Thin data hurts most. When a team had fewer than three matches of data, the model scored 1.163, worse than guessing. With three or more it scored 1.038.
Over 3,519 matches the bookmaker odds were a better forecast than our model in every group. That is why we show both numbers side by side and never label a pick as value.
Calibration: does 70% mean 70%?
A model is well calibrated when its 70% picks win about 70% of the time. We grouped every settled live prediction by the probability the model gave its pick, and counted how often that pick actually came in:
| Model said | Matches (old v3.1) | Came in (old) | Matches (new v4) | Came in (v4) |
|---|---|---|---|---|
| 30–40% | 16,548 | 39.0% | 679 | 37.3% |
| 40–50% | 26,532 | 41.0% | 667 | 45.0% |
| 50–60% | 14,085 | 45.1% | 399 | 54.4% |
| 60–70% | 6,532 | 52.4% | 147 | 64.6% |
| 70–80% | 3,054 | 59.4% | 81 | 71.6% |
| 80–90% | 1,312 | 67.6% | 35 | 80.0% |
| 90–100% | 334 | 70.7% | – | – |
Source: 68,404 settled live predictions, 23 March to 11 October 2026. The old model ran on all competitions; v4 runs only where the data is best, so the two columns are not the same matches.
The old model was over-confident, and more so the surer it was: at about 84% its pick came in 67.6% of the time, at about 94% only 70.7%. Our over/under model shows the same pattern: when it said about 74%, the line came in 57.4% of the time over 18,717 matches. The newer v4 model is much closer to the diagonal, but on far fewer matches.
The old model’s high numbers were too high. Where it said 74%, reality was 59%. That is the main reason we dropped the “AI confidence” labels and now show the market probability next to every model number.
Old model against new model, on the same matches
The calibration table is not a fair fight: v4 only runs in the competitions with the best data. So we also compared both models on exactly the same 1,639 matches:
| On the same matches | Old model (v3.1) | New model (v4) |
|---|---|---|
| Matches | 1,639 | 1,639 |
| Picked the right result | 47.3% | 47.3% |
| Log-loss (lower is better) | 1.217 | 1.032 |
Source: Daily Sport Pick live predictions, April to October 2026, matches with probabilities from both models.
Both models picked the right result equally often, 47.3%. The difference is in the probabilities: v4 scores 1.032, the old model 1.217. For scale, always predicting the average scored about 1.08 in our September test, and the bookmaker odds about 1.00. The gap in log-loss is larger than the gap in the calibration table, which points to confident mistakes on the outcomes the old model did not pick: log-loss punishes a 3% chance that happens far harder than a 60% pick that misses.
The new model does not pick more winners. It is more honest about how likely they are. On this set it beats guessing clearly; whether it beats the bookmaker market we will only claim once we have measured it on fresh matches.
What happens when the models agree
For each match we count how many of the three models favour the same outcome:
| Models agreeing | Matches | Model said | Came in |
|---|---|---|---|
| One model | 31,737 | 47.6% | 40.2% |
| Two models | 31,789 | 50.6% | 47.5% |
| All three | 1,315 | 54.4% | 50.0% |
Source: settled live predictions, March to October 2026.
The system also gives each match an internal ensemble score from 0 to 100. It ranks matches well: below 10 the favoured result came in 35.7% of the time, above 90 it came in 76.1%, and the 50 matches scoring 100 came in 84.0% of the time. It is a ranking number, not a probability, and it does not decide which picks we publish.
Agreement helps: 50.0% when all three agree, 40.2% when only one model backs the outcome. But even then the model said 54.4%, so agreement does not cure over-confidence. Only 6 of our paper picks had full agreement; 5 won, far too few to build on.
The three models
- Machine learning. A gradient-boosted decision-tree model trained on 28,000+ historical matches. The newer v4 version handles the competitions with the best data; the older v3.1 acts as a fallback elsewhere.
- Poisson. Estimates how many goals each team is likely to score from attack and defence ratings per competition, corrected for home advantage, and turns those expected goals into a full score matrix.
- Dixon-Coles. Extends Poisson with a correction for low scores such as 0-0 and 1-1 (Dixon & Coles, 1997), fits by maximum likelihood, and gives recent matches more weight. Fitted per competition.
- The ensemble. Where the machine-learning model has reliable data, the three are weighted 60% machine learning, 20% Poisson and 20% Dixon-Coles. Elsewhere Poisson and Dixon-Coles share the weight. A model without reliable data for a match is left out and the weights are re-normalised.
The machine-learning model uses features such as:
- ELO ratings, global plus separate home and away ratings
- Recent form of both teams
- Goals and expected goals, scored and conceded per game
- Number of prior matches, so the model knows how much data it has
Every feature is computed point-in-time: only information available before kick-off is used. The weak spot is early in a season and in cup competitions, when a team has played few matches in that competition.
How we choose which picks to publish
Since 18 September 2026 the same rules run every day, with no manual selection:
- Competitions: Premier League, La Liga, Serie A, Bundesliga, Ligue 1, Eredivisie, Primeira Liga, Belgian Pro League, Championship, Brasileirão and MLS.
- Enough data: both teams need at least three prior matches of data.
- Accumulator picks: the highest model probability, at least 70%, at odds of at least 1.30, and no more than 18 percentage points away from the market.
- Extra picks: model probability of at least 50%, odds between 1.30 and 4.00, no match-result market, and no more than 10 percentage points above the market. Ranked by probability.
- Cross-division check: fixtures where a team’s data comes from a different division are skipped.
- No fallback picks: if nothing meets the rules, we publish fewer picks, or none.
Our published tips before and after the new rules:
| Period | Main tips | Won | Return | Extra tips | Return |
|---|---|---|---|---|---|
| 10 Mar to 17 Sep | 448 | 64.5% | −4.7% | 225 | −21.5% |
| 18 Sep to 11 Oct | 29 | 75.9% | +7.7% | 18 | −10.7% |
Source: Daily Sport Pick published tips, 1 unit per tip, settled at the odds published. Full record on our results page.
The pick tables tell a more sober story. Since 18 September the accumulator table’s model said 77.8% on average, 66.2% came in, and the odds implied 74.2%. On those picks the market was still closer to the truth than we were.
The first 29 main tips under the new rules returned +7.7%, but 29 tips prove nothing: our monthly record has swung between −22.4% and +13.5%. We will publish a full new measurement once there are several hundred.
What you see on the site
- Model % and market %. On picks, analysis pages and market tables, our model probability sits next to the probability implied by the odds. Market % removes the bookmaker margin: odds of 1.80 and 2.10 on two outcomes become roughly 54% and 46%.
- Market view elsewhere. For competitions without picks, analysis pages show only the market probabilities.
- No value labels. A gap between model and market is not an edge. Our value betting test shows why.
- Score heatmap. The grid on each analysis page shows every scoreline from 0-0 to 4-4: from our model’s expected goals where we publish picks, from the bookmaker odds elsewhere.
- Check it yourself. Our AI model statistics page shows the ranking score against settled results, and we published a postmortem of three bugs in our closing line tracker, including what we still cannot measure.
Frequently asked questions
What prediction models does Daily Sport Pick use?
Three: a machine-learning model (gradient-boosted decision trees) trained on 28,000+ historical matches, a Poisson goals model and a Dixon-Coles model. Where the machine-learning model has reliable data they are weighted 60/20/20; elsewhere Poisson and Dixon-Coles share the weight.
Does your model beat the bookmakers?
Not so far. On 3,519 matches between April and September 2026 the probabilities implied by bookmaker odds were a better forecast than our model in every market and competition group we measured. Our newer v4 model is clearly better than the old one, but we have not yet shown that it beats the market.
How accurate is the AI model?
On 1,639 matches where both versions ran, the old and the new model each picked the correct match result 47.3% of the time, against 33% for a blind guess. Across all 68,000+ live predictions since March 2026, which include weaker competitions, the figure was 43.9%.
Is the model well calibrated?
The old model was not: when it said about 74%, its pick came in 59.4% of the time. The new v4 model is much closer: when it said about 74%, its pick came in 71.6% of the time, on fewer matches.
What happens when all three models agree?
Their pick comes in more often: 50.0% over 1,315 matches, against 40.2% when only one model backs the outcome. But the models said 54.4% on those matches, so agreement does not remove over-confidence.
What is the market probability?
The probability implied by the bookmaker odds, with the bookmaker margin removed. Odds of 1.80 and 2.10 on two outcomes become roughly 54% and 46%.
Get free tips, every day
Daily AI-backed football tips in your inbox. Free, no spam, unsubscribe anytime.
18+ only. Bet responsibly. Gambling can be addictive.