How OddsFlow AI Works — The Dixon-Coles Model Explained
Ever wondered how a computer can predict football match outcomes better than your most passionate fan friend? No black boxes, no hype, no "trust us" promises. This is a step-by-step explanation of the statistical model behind OddsFlow's predictions, with practical examples you can follow and verify yourself.
What Is the Dixon-Coles Model?
Published in 1997 by statisticians Mark Dixon and Stuart Coles, the Dixon-Coles model remains the gold standard for football match prediction in academic statistics. It is a Bivariate Poisson model — which sounds intimidating but works on a simple idea: if you know how many goals a team tends to score and how many it tends to concede, you can calculate the probability of every possible scoreline.
The model estimates four key components for every match:
- 1Attack strength — how dangerous a team is going forward
- 2Defensive strength — how difficult it is to score against them
- 3Home advantage — the measurable boost from playing at home
- 4Low-score correction — a factor that adjusts for the fact that 0-0, 1-0, 0-1, and 1-1 scorelines occur more frequently than standard Poisson predicts
With these four ingredients, the model generates probabilities for every possible scoreline (0-0, 1-0, 2-1, 3-2, and so on), which are then summed to produce match outcome probabilities, Asian Handicap lines, and Over/Under totals.
Why Poisson?
Because goals in football naturally follow a Poisson distribution. A team averaging 1.5 goals per match does not score exactly 1.5 every time — sometimes 0, sometimes 3, rarely 5. The Poisson distribution models exactly this kind of variability.
Liga MX Example: Club América vs Puebla
Let's walk through a real scenario. The model calculates these parameters for an América vs Puebla match at Estadio Azteca:
| Parameter | Club América (Home) | Puebla (Away) |
|---|
| Attack strength | 1.42 | 0.88 |
|---|---|---|
| Defensive strength | 0.82 | 1.25 |
| Home advantage | 1.18 | — |
The model then runs these expected goals through a Poisson distribution to generate scoreline probabilities:
| Scoreline | Probability |
|---|
| 1-0 | 14.8% |
| 2-0 | 13.2% |
| 2-1 | 10.1% |
| 1-1 | 7.6% |
| 3-0 | 7.4% |
Example 2: Cruz Azul vs Chivas — Draw Probability
Derbies are different. When two strong teams meet, draw probability rises. For Cruz Azul vs Chivas at Estadio Azul:
- Cruz Azul expected goals: 1.40
- Chivas expected goals: 1.04
- Result: Cruz Azul 45.8%, Draw 26.1%, Chivas 28.1%
That 26.1% draw probability is key. If the bookmaker's odds imply only 22%, that is a value bet — exactly what OddsFlow hunts for.
How Probabilities Become Odds
The conversion is straightforward:
- Decimal odds = 1 / probability (71.2% = 1.40 decimal)
- American odds: Below 2.00 decimal = negative (1.40 = -250). Above 2.00 = positive (3.83 = +283)
Where Value Exists
Suppose the model calculates Tigres at 58% to beat Monterrey in the Clasico Regio. Fair odds: -138. If your book offers -120, you are getting a better price than fair — that is a value bet. If they offer -165, the AI says pass.
Why We Don't Use ChatGPT
This is the most common question we receive. The answer is definitive: ChatGPT is not designed to predict sports outcomes.
ChatGPT is a language model. It cannot access real-time odds, cannot run Poisson distributions, and does not update its parameters from match results. When it says "65% chance," that number comes from text patterns, not mathematical calculation. It hallucinates confidence.
OddsFlow operates differently at every level:
| Capability | OddsFlow | ChatGPT |
|---|
| Real-time odds | Scans 10+ bookmakers every 10-20 seconds | No access |
|---|---|---|
| Statistical model | Dixon-Coles + 10,000 Monte Carlo sims | None |
| De-vigging | Shin Method margin removal | Cannot |
| Learning from results | Updates parameters after every match | Does not |
| Explainability | Full parameter audit trail | "Trust me" |
Closing Line Value — How to Verify It Works
Any prediction platform can show you winning screenshots. What separates real systems from noise is Closing Line Value (CLV).
CLV measures whether your entry odds were better than the closing line (the odds at kickoff). If you consistently beat the close, the model has real edge — the market eventually arrived where the model already was.
Example: OddsFlow flags Club América -1.5 at +130 on Monday. By Friday kickoff, the line has moved to +105. That positive CLV means the market confirmed the model's assessment.
Why CLV matters more than win rate: you can win 60% of bets and lose money with bad odds. You can win 48% and profit with excellent odds. CLV tells you whether your prices are better than the market's — the only thing that matters long-term.
OddsFlow publishes CLV data on the verification page for full transparency.
Time Decay — Recent Matches Matter More
Not all historical matches carry equal weight. The model applies exponential time decay:
| Recency | Approximate Weight |
|---|
| Last 30 days | 100% |
| 1-3 months | 75% |
| 3-6 months | 50% |
| 6-12 months | 25% |
| Over 1 year | 10% |
The 5-Step OddsFlow Pipeline
- 1Odds collection — scan 10+ bookmakers every 10-20 seconds for market consensus
- 2Shin Method de-vigging — strip bookmaker margins to reveal true probabilities
- 3Dixon-Coles parameters — calculate attack strength, defensive strength, home advantage, and league coefficients
- 410,000 Monte Carlo simulations — generate random scorelines per match using Poisson distributions
- 5Signal generation — flag value bets only when the model edge exceeds a minimum threshold across multiple bookmakers
Start Free — No Card Required
See OddsFlow predictions in action across Liga MX, Premier League, La Liga, and 19+ leagues.
Create your free account — no credit card needed. Verify the results yourself and decide if the model delivers value.
We do not promise to make you rich. We promise to show you exactly how our model works, with verifiable results and total transparency.