Can we really trust AI predictions in football? 5 overlooked risk signals from a Bayesian model

The Model Doesn’t Speak, But the People Do

I built a Bayesian model to predict Premier League outcomes—not to replace instinct, but to amplify it. Every probability curve I traced was whispering: ‘What if the data is wrong?’ Not because the numbers are flawed—but because we forgot who reads them.

The 24 Who Were Never Invited

There were 24 analysts at our firm. None were invited to the final review. Not because they lacked skill—but because their voices didn’t fit the algorithm’s rhythm. We optimized for win rates, not for wisdom.

Small Green House, No One Saw It

They called it ‘Small Green House’—a quiet corner of our server where raw decisions got stored. No one took photos there. It wasn’t about aesthetics—it was about accountability. A place where logic sleeps, and no one dares to wake it.

The Algorithm That Forgot Its Name

Our system learned from match data, but never learned its own name. It predicted wins based on priors—not on players’ souls. When the model whispered ‘Should we trust this?’, no one answered.

The Quiet Tension Between Code and Humanity

I grew up in Croydon: mother a Nigerian nurse, father a Scottish engineer. I speak in equations—but I listen in silence. When you ask if AI can replace judgment… don’t look for answers in code.

Five Risk Signals You Didn’t Hear

1. Overfitting bias masked as precision
2. Training data without ethical review
3. Win rate obsession over fairness
4. Ignoring human context in model design
5. Silent rejection of interpretability