About KayLoL

This is the thing worth understanding before you judge the tool, because it is the one place KayLoL deliberately makes itself look worse.

A draft-prediction model can be evaluated two ways. The easy way is a random split: shuffle all your games, train on most, test on the rest. The hard way is a held-out patch: train only on games that happened before a patch, then test on the patch itself — the way the model will actually be used, on a meta it has never seen.

Random splits produce better-looking numbers, because the test games sit in the same meta as the training games and some of the answer leaks across. Filtering to only the highest division, or discarding games that ended early, flatters them further. KayLoL is measured the hard way, on a whole unseen patch, across the whole ladder. That costs several points of headline accuracy and it is the honest number.

It also means the number to compare is not the one most sites lead with. Accuracy publishes three separate layers of evidence, ordered by how hard each is to fake, and never merges them into a single figure. The hardest to fake is the predictions this API returned before the games were played, which cannot be reconstructed after the fact; that layer is still being collected, and its numbers appear once enough of those games have been scored.

Most tools optimize for telling winners from losers. KayLoL also cares whether 57% actually means 57% — that across every draft it called at 57%, blue really does win about 57 of 100. A model can rank drafts well and still be badly calibrated, and a confidently wrong probability is worse than a hedged one when you are using it to make a pick.

That is why the metrics page leads with a calibration curve rather than an accuracy score, and why the site keeps saying the same uncomfortable thing: a draft is a small part of a game. Expect probabilities near 50%, and read the direction rather than the last decimal.

Completed ranked games collected from Riot’s public Match API — the same games Riot already publishes to any client, and the same source every match-history site is built on. They train the model and produce the champion statistics behind the Tier list, and they also make up each account’s record — games, results and champions — that the pages which take a Riot ID read from.

Pages that take a Riot ID show public match data for whichever account is entered — depending on the page, its rank, its record on each champion, its recent games, or on head to head the games two accounts played in the same match — as other League stats sites do. The site cannot check that an account belongs to the person who typed it. What is stored, what is deliberately never stored, who else handles any of it, and what can be removed on request are all in the Privacy Policy.