Football prediction is a signal-combination problem. No single source — not odds markets, not historical stats, not injury news — is reliable enough on its own. The best predictions blend multiple independent signals and measure how strongly they agree. Here is exactly how Sportdico does it.
The three signals
1. Market odds (highest weight)
Bookmaker odds are not prices — they are crowd-sourced probability estimates, adjusted for margin. When millions of pounds flow onto one outcome, the market is usually right. We strip the bookmaker's overround (typically 5–8 %) using the standard normalisation method to recover the implied probability of each outcome.
For a typical match this gives us three numbers that sum to 100 %:
| Outcome | Raw odds | Implied prob (with margin) | Vig-removed prob |
|---|---|---|---|
| Home win | 2.10 | 47.6 % | 50.5 % |
| Draw | 3.40 | 29.4 % | 31.2 % |
| Away win | 3.80 | 26.3 % | 27.9 % |
| Total | — | 103.3 % | 100 % |
The market signal gets the highest weight because it aggregates the most information.
2. Statistical model (Dixon-Coles Poisson)
We model each team's attacking and defensive strength as a pair of Poisson parameters, updated weekly from recent results. For each match we simulate the full scoreline probability matrix — every possible score from 0–0 to 6–6 — and sum across outcomes to get home win / draw / away win probabilities.
The model also quantifies:
- BTTS probability — P(home goals ≥ 1) × P(away goals ≥ 1)
- Over 2.5 probability — P(total goals ≥ 3)
Dixon-Coles adds an adjustment for low-scoring draws (0–0 and 1–1), which are systematically underestimated by a naïve Poisson model.
3. Editorial review
For selected high-stakes matches — cup finals, relegation deciders, derbies — our analysts add context the model cannot read: confirmed starting XIs, injury to a key goalkeeper, extreme weather, recent travel schedule. This adjusts the published confidence score up or down and can override the model's market pick.
How confidence scores are calculated
The confidence score (0–100) measures agreement between the market signal and the statistical model on the same outcome. Two inputs:
- Signal agreement — how close are the two probability estimates for the top outcome? Close = higher confidence.
- Outcome dominance — how much does the favourite lead the second-most-likely outcome? A 65 % favourite scores higher than a 40 % favourite.
Scores above 70 are reserved for matches where both signals point clearly to the same outcome. Below 50, we publish the prediction but flag it as speculative.
What we predict (and what we do not)
We publish one primary market per match: the 1X2 result, or BTTS / Over 2.5 when those markets offer a clearer edge. We do not publish correct score tips from the model alone — scoreline probabilities decay quickly and are extremely sensitive to lineup changes.
Frequently asked questions
How far in advance are predictions published?
Typically 24–48 hours before kickoff. Predictions for matches more than 48 hours away are not published because form data and injury news is too stale to be useful.
What happens if a match is postponed?
The prediction is unpublished automatically and reissued if the match is rescheduled within the current season.
Does the model learn from its own predictions?
No. The model is trained on match results, not on previous predictions. This avoids feedback loops where a confident wrong prediction would poison future training data.
Are predictions guaranteed to be correct?
No prediction is guaranteed. Even a 90 % confidence score means roughly one in ten tips will be wrong. Use predictions as one input among many, not as the sole basis for a betting decision.