I built a Poisson-based football probability model with a human-context layer — looking for critical feedback
Hi everyone,
I'm a data scientist and I've been working on a football probability tool called TODO Mining Predictor.
The core model uses Poisson distributions to estimate scoring rates (λ) and generate a full score probability matrix. From that distribution, the system calculates probabilities for outcomes such as:
- Home / Draw / Away
- BTTS
- Over / Under
- Double Chance
- Most probable scorelines
But there's a part of the approach I'd particularly like feedback on.
Instead of treating the statistical model as the final answer, I added an optional Guided Human Judgment (GHJ) layer.
The idea is that historical data may not fully represent the conditions of the upcoming match. The analyst can therefore incorporate contextual factors such as tactical matchup, fatigue, rotations, offensive/defensive absences, goalkeeper form and expected match tempo.
These inputs modify the scenario, and the probability distribution is recalculated.
I'm deliberately not posting accuracy claims or a win rate, because I don't think a handful of successful predictions would demonstrate that the model has an edge.
What I'm interested in now is the next step: evaluating whether these probabilities contain useful information relative to market prices.
For those of you who build or evaluate betting models:
What evidence would you want to see before considering a model like this useful?
For example:
- Calibration?
- Brier score / log loss?
- Closing Line Value?
- Comparison against no-vig market probabilities?
- Out-of-sample backtesting?
- ROI over a sufficiently large sample?
- Performance with and without the human-context layer?
I'm especially interested in criticism of the methodology.
I can also share a real match example where the model's highest-probability outcome did not occur, including the inputs, λ values, probability distribution and what happened afterward. I think a failed prediction may be more useful for discussion than cherry-picking a winner.
I'm trying to determine what should be validated before adding more features to the tool.
Thanks — I'd genuinely appreciate technical feedback.