Every prediction is frozen before kickoff and scored after the final whistle. Only the base model is scored: the agent's adjustments depend on news you bring, and there is no honest way to grade them against history.
Predictions for every fixture are frozen 48 hours before kickoff, before any team news, and can't be edited after that. Once a match has been played and its result is in, it is scored and shows up here, usually the next morning.
Until then, the backtest shows how the same model did on 1,827 past matches.
By kickoff, the closing line has priced in everything public, including news our goals-only model never sees. That is why the model on its own sits just behind it.
The window is before the odds settle. Lines open early in the week and keep moving until kickoff. Bring the agent a lineup, an injury or a forecast before the market has caught up, and it reprices the match in seconds.
We check every one of those chats against where the market actually finished, so you can see when the news put you ahead.
Each row is a chat that brought news before kickoff. Points are the change from the opening line for the outcome the news moved most. A green circle means the agent's odds landed closer to the closing line than the opening line did, in the direction the market later moved. This is a separate measure from the model's score above, and chats stay private: only the odds are shown.