What the win probability on a CS2 prediction means, what it does not mean, and how to use it.
A call of 68% says that across matches that look like this one, the picked side comes out ahead roughly two times in three. It does not say this particular match is settled. Close to a third of calls at that number are expected to be wrong, and that is not a flaw in the number — it is what the number said.
The only way to know whether the percentages mean anything is to check them against results, in bulk, in public. Our published record is that check.
Two evenly matched rosters produce a number near the middle, because that is the honest answer. Counter-Strike is played to a small number of rounds on maps that both sides have prepared, and one team having a better month does not make the other incapable of winning a best-of-three.
A prediction that reads 95% is describing a mismatch, not confidence. Those exist, and they are less interesting than they look: everyone can see them, so nothing about them is worth much.
Recent results against comparable opposition, the map pool each side is likely to reach, roster changes that have actually been played rather than announced, and whether the match is online or on stage. Each of those shifts the estimate by a few points; none of them decides it alone.
The picked side is simply the side above 50%. Reading the pick without the number throws away most of what was said: "we favour this team" is a very different claim at 55% than at 80%, and the page states both because only the pair is useful.
Longer series favour the stronger side, because a single bad map matters less. The same two teams will produce a flatter number in a best-of-one than in a best-of-three, and the format is on every prediction page for that reason.
Every prediction lists what could make it wrong. That section is not a hedge; it is the part that tells you which assumptions the number rests on. When a risk listed there is the thing that actually happens, the call was wrong for a reason that was written down in advance — which is the only kind of wrong worth anything.
A single call is one sample. Judged on its own, a correct one at 55% looks like insight and an incorrect one at 80% looks like failure, and neither reading survives contact with the next fifty matches. The archive exists so the calls can be read together, including the ones that missed.
Each prediction also carries a confidence level. It is a band the probability falls into, shown so a page can be scanned without reading every figure. It never disagrees with the percentage, because it is derived from it — what the levels mean covers how the bands are drawn.