A complex neural network using 150 years of match data on a GPU failed because the historical data was irrelevant. A much simpler model, running on a laptop, proved more accurate by focusing only on two recent signals: current squad market value and in-tournament form. This demonstrates that for some prediction tasks, data relevance and model simplicity trump data volume and computational power.
The AI model's advantage came from its ability to solve the numbers problem: objectively assessing team strength based on squad value and form. The human journalist's correct predictions, however, came from reading intangible factors like team morale, a coach's influence, or player behavior under pressure—elements a Poisson model cannot see. The true takeaway is that AI and human intuition are suited for different parts of the same problem.
Instead of just comparing the human and AI's picks, a 'hedge' algorithm treated them as two competing experts. It created a combined consensus pick for each match by assigning dynamic weights based on their recent performance. This system automatically leaned on whichever expert—human or machine—had proven more reliable lately, creating a robust, self-correcting prediction engine.