Why a CS2 prediction carries a confidence level as well as a probability, and how the levels are drawn.
Every prediction has a win probability. The confidence level is the band that probability lands in — a label over a range, not a second opinion arrived at separately. It can never disagree with the number, because it is derived from it.
A page of raw percentages is hard to scan and easy to over-read. Sixty-two and sixty-four are not meaningfully different calls, and shown side by side as numbers they invite a comparison that the underlying estimate cannot support. A band says "these two are the same kind of call", which is the truthful reading.
Where one band ends and the next begins is configured rather than hard-coded, and it has been adjusted as the record grew. That matters for reading the archive: a call from months ago carries the band it was reviewed under, so the historical numbers stay comparable instead of being quietly re-labelled every time the scale changes.
It is a claim about a larger share of similar matches. The top band is expected to be wrong sometimes, and when it is, that is a normal outcome rather than a failure of the level. How the record is measured explains what "expected" means here.
A call in the lowest band is saying the match is close to even. That is worth publishing: knowing a match is a coin flip is more useful than a confident-sounding sentence about it, and pretending otherwise is how a record gets padded with claims nobody checked.
The level appears as a rating and its name together. Colour carries no meaning on its own anywhere on the site, so the level is readable without seeing colour at all.
The interesting question about bands is whether the higher ones actually land more often. The accuracy page answers it directly, per level, from the same reviews that produce the overall number — so a band that is not earning its label has nowhere to hide.
It is not a stake size, a recommendation, or a rating of the teams. It describes how far from even one estimate sits, and nothing else. What a prediction is and is not covers the distinction in full.