Bookmakers vs the model
Nine Eurovision finals replayed, country by country: what the betting market expected, what this model expected, and what really happened.
The methods side by side
All three were given the same nine finals (2017 to 2026, no contest in 2020) without seeing the votes. Each number is the average across those finals.
Every finalist, every method
Rows run in the real finishing order. Each method's column shows its chance to win and the place it predicted. Gold means it was within two places; pink means it was off by seven or more. Years marked with a dot are the ones the favorite lost.
Semi-finals
Who gets through
The same replays for the semi-finals, made the way today's board is: no songs, no odds, only the real line-up, the voting countries, recent results and voting habits. Gold outline: in the model's top ten. Green edge: really qualified. Struck through: went home.
What the upsets teach us
How the model should get better
Each idea is only kept if it makes the replays more accurate. With nine finals to test on, small tweaks can look good by luck, so every test is shown here, including the ones that failed. Log loss measures how surprised a method was by the real winner: lower is better.