The secondary signal and confidence

Two more pieces sit around the pick: an outside opinion and a confidence rating. Both are about honesty regarding how much to trust a given prediction.

API-Football's own prediction

As a cheap cross-check, the model can pull API-Football's own /predictions for a fixture — its suggested winner, a piece of advice, and home/draw/away percentages — and show it as a secondary opinion beneath the model's pick. When the two agree, that's mild reassurance; when they disagree, it's a flag to look closer.

To protect the request budget, this outside prediction is fetched only for the fallback fixtures (the ones without local history, like World Cup matches), not for every club game. It's a supporting signal, never the deciding one.

Confidence levels

Each prediction carries a confidence level — High, Medium or Low — and it's driven mainly by how much data stood behind it. A match with a full last-10 window for both teams and a real head-to-head history, computed from stored local history, can reach High. Thin samples, or predictions leaning on the API fallback, land at Medium or Low.

Confidence here is about the foundations of the estimate, not a promise about the outcome. A High-confidence pick still loses sometimes; it just means the model wasn't guessing in the dark.

Confidence measures how solid the inputs were, not how sure the result is. More data, higher confidence — but never certainty.