By Reece Bentham Finance, aspiring accountant, Yorkshireman through and through
I should probably start with a confession.
Some of these predictions will be wrong.
There. I’ve said it before anyone screenshots the bracket and waits patiently for football to make fools of us all.
As a finance person, I am used to forecasts being challenged. As a Spurs fan, I am used to optimism being punished. So I feel strangely qualified to write this blog.
The model has now produced its first view of the World Cup. Not a prophecy. Not a guarantee. Just a view based on the evidence available at the time. That distinction matters because the purpose of this project is not to pretend we have solved football. We have not. Football is low-scoring, emotionally unstable and occasionally decided by somebody’s shin in the 89th minute.
Why we built the dashboard this way
What we are trying to demonstrate is something much simpler: what responsible predictive analytics looks like when it is exposed to reality.
Most predictive projects quietly move on when the first version gets something wrong. We wanted to do the opposite. We wanted to publish the prediction, make the assumptions visible, compare it with what actually happens and then improve the model in public. That is why the World Cup dashboard site exists. It is not simply a place to publish a bracket. It is a place to show how a prediction changes as new evidence arrives.
The first version of the model, V1, was built around PI-TSI, our team strength measure derived from recent form, goals scored, goals conceded and the relationships identified through Panintelligence Analytics. It gave us a transparent starting point and a bracket that anyone could challenge.
Reality challenged the first version
Then the tournament started.
After the first 20 completed matches, V1 had correctly predicted 8 winners. That is 40% accuracy. Not disastrous. Also not particularly impressive. More importantly, it gave us something we had not previously possessed: evidence. For the first time, we could compare the model against reality and start asking better questions. Was the model failing randomly? Was it consistently underestimating something? Had we missed an important signal?
V1 Predictions vs Reality
As we reviewed the results, a pattern began to emerge. The model appeared to be missing information that many football fans instinctively consider important: the relative strength of the teams before the tournament began.
That led us to build V2.
What the model learned
The key principle was that the World Cup results themselves could not become part of the training data. If the model learns from the matches it is supposed to predict, the exercise becomes meaningless. Instead, we froze the tournament results and introduced one additional pre-tournament signal: FIFA ranking points difference.
When we rebuilt the decision tree, the result surprised us. FIFA points difference became the dominant predictor. That does not mean form, attack and defence suddenly stopped mattering. It means the strongest missing signal turned out to be the gap between how strong teams were already considered before the tournament started.
The important thing is that we can explain the change. V1 correctly predicted 8 winners from the first 20 matches, while V2 predicted 11. Accuracy increased from 40% to 55%.
The model improved, not because we changed it until we liked the answers, but because reality showed a weakness and the evidence pointed us towards a better approach.
V2 Predictions vs Reality
The value of keeping both versions visible is that improvement becomes measurable. Rather than quietly replacing one model with another, we can explain what changed, why it changed and whether it actually performed better.
England gave us a useful example
England’s recent match against Croatia provided a good example of why evaluating a model is more complicated than simply counting wins and losses. Before the game, V1 predicted a 1-0 England victory and V2 predicted a 2-0 victory. England eventually won 4-2.
On paper, both models got the winner right and the score wrong.
Whether that counts as success depends on what question you are asking.
For tournament simulation, identifying the likely winner matters enormously because it affects points, qualification routes and knockout progression. For score prediction, however, both versions clearly missed something about how the match would unfold. The result is useful precisely because it sits in the uncomfortable middle ground. The models were not completely right, and they were not completely wrong. The match itself helps explain why.
A final scoreline of 4-2 makes it look as though England were comfortably in control. They were not. Croatia equalised twice, the teams went into half-time level at 2-2 and the game remained far more competitive than either model anticipated. England eventually pulled away through goals from Jude Bellingham and Marcus Rashford, but the route to the result was far less comfortable than a 4-2 scoreline suggests.
That is exactly the sort of outcome predictive models struggle with. The models correctly identified England as the stronger side, but they could not account for the twists inside the match itself: momentum swings, defensive mistakes, tactical adjustments and moments of individual brilliance. Football rarely unfolds in a straight line, even when the result appears to justify the prediction.
That is exactly the kind of feedback we want.
What the model still doesn’t know
The current model is still operating with incomplete information. It does not read breaking team news. It does not know who is carrying an injury. It does not know when a manager is about to abandon Plan A after twenty minutes. It does not know whether a goalkeeper is about to produce the performance of their career. And, apparently, it still does not contain a variable for England’s kit going missing at a crucial moment.
Football has a long-standing commitment to chaos. Models have a long-standing commitment to evidence. The tension between those two things is what makes this project interesting.
As more matches are played, we will continue testing whether the assumptions still hold, whether new signals deserve to be included and whether future versions genuinely outperform the previous ones. The objective is not to keep modifying the model until it tells us what we want to hear. Tempting though that may be, depending on what it says about England. The objective is to make the next prediction more informed than the last.
Why we keep every version
That is why V1 remains visible. V2 now sits beside it and future versions will follow the same pattern. The goal is not to replace history every time we learn something new. The goal is to measure improvement. If a new signal helps, we should be able to explain why. If a model performs better, we should be able to show the evidence. Keeping previous versions visible makes that possible.
The phrase we keep returning to throughout this project is simple: the data will change, the model will learn and the predictions will move. That is not a slogan. It is the entire point.
The interesting question is not whether the first bracket was perfect. It clearly was not. The interesting question is whether the next version becomes better because reality challenged the last one.
Football gets the next word
The model has made its call.
Football gets the next word.
And if the data starts moving England in the right direction, I will be treating that as a sign of model maturity.
Not bias.
Definitely not bias.
