Accuracy

One bundle, five models, each trained for a different amount of knowledge about the lobby. They are not quality tiers: sending a board to the wrong one gives a confident number about a situation nobody is in, so the site picks the model from what you have told it — and every table below names which one each number belongs to.

Official competitive tournament maps scored walk-forward with zero post-game leakage, evaluated against actual map outcomes (bwin) on the same 5 metrics — no odds or betting.

Tournament maps scored

2,117official tournament maps

Post-Draft (Full composition)

0.7325AUC

Pre-Draft (Rosters alone)

0.7256AUC

Draft Lift

+0.0069post minus pre

When we said 62% about a tournament map, how often did blue win?

2,117 official maps, both arms on the same games. This is a far smaller sample than the solo-queue curve above and the buckets show it — read the interval, not the dot. Horizontal: what we predicted. Vertical: what happened. A perfectly calibrated model sits on the dashed line. Each dot is one probability bucket, sized by how many predictions fell in it; the bar is its 95% interval. Wide bars mean not enough games yet — not a wide model.

Put in the same units as the rest of the page: Post-Draft (Full composition) has a Brier score of 0.2049, and miscalibration accounts for 0.001319 of it — about 0.644%. Nearly all of the error is the coin the game flips and what the model cannot separate, not the probability being wrong. Overall we under-call blue by 0.99%, a direction an average gap cannot tell you because it throws the sign away.

How predictive accuracy improves when moving from pre-draft roster ratings alone to the full 10-champion composition and draft state.

Both rows score the same maps, so the difference between them is the draft and nothing else.

Our own measurement, on a split we chose. The number is ordinary; the split is the part worth checking.

Split

patch holdoutno game from the test window was in training

Measured on

2026-10-06the patch build these numbers come from

Serving now

2026-10-20a daily refit of the same recipe, on everything since

Where the test games came from

Sized by real days. The two windows meet at Sep 22 and do not overlap.

Trained on

4,952,043 gamespatches 1618 · Sep 9 to Sep 23 · 14 days

Held out and tested on

5,385,379 gamespatches 1619 · Sep 22 to Oct 6 · 14 days

Why training holds so few patches: KR and EUW began collecting on 31 Aug 2026, and one model serves all three regions, so it trains only on days all three have. A build also keeps its newest patch back to test on, which leaves 1 of the 4 patches a build normally trains on. Each new patch adds one until the window is full. Player histories in KR and EUW are just as young, so the modes that know the players have the most to gain; the patch builds below will show whether they do.

What a random split would look like instead

An illustration, not our data. Published figures elsewhere come from splits shaped like this one: the same patches on both sides, so the model is tested on the meta it trained on. Same amount of test data, drawn from inside the training period rather than after it.

What each mode scores

The held-out patch above. The axis starts at 0.50, which is a coin flip — not at zero, because no model lives below it.

The same measurement, remade every patch

Each build re-selects its features and re-fits from scratch, then scores a patch it has never seen. Every facet is 0.080 AUC tall, so the same rise means the same thing in all five — but each is centerd on its own mode, because these are five different questions and not a ranking. Most of them barely move, and that is the finding rather than a flat chart: across 5 rebuilds, every mode but Mid-draft, solo queue stays inside 0.044 AUC end to end.

The first two builds hold out the same patch over a growing window (516,665 games, then 1,092,703), so they are not independent points. A fall is not automatically a regression. The 26 Aug 2026 build is the one that stands out here, and it is a correction: Mid-draft, solo queue lost cumulative identity visibility there, which had been showing it the histories of everyone who had already picked — information nobody has in a real solo-queue champion select.

Points from builds whose numbers include a known leak (described under the chart above): Full draft, players known (4/5), Mid-draft, roster known (4/5), Mid-draft, solo queue (4/5).

The spread between modes is the product, not a defect. Anonymous modes know only the champions; the identity modes also know who is playing them, which is worth more than any single feature we have found — and it is the thing a champions-only tool cannot add.

Every ranked game of a day, scored by a model frozen before that day happened. No split to choose: the future is the test set.

The same day of every patch, since we started

One facet per mode, 0.0400 AUC tall. The axis is days since the patch shipped, so every point averages that position across patches. A day’s standard error, about 0.002 AUC at 80,000 games, is roughly a tenth of a facet, and a single wobble that size is noise rather than news. The dashed line is that mode’s held-out-patch AUC from the Oct 6, 2026 build — a different measurement on different games, which is exactly why it is worth seeing whether the days sit on it.

When we said 62%, how often did blue actually win?

10,762,569 predictions over 28 days (Sep 23, 2026 to Oct 20, 2026), one facet per mode — the same games as the figure above, grouped into five-point ranges of what we forecast rather than by when. The solid line is fitted to the buckets, weighted by how many predictions each holds; the dashed one is where a perfect forecast would sit. Horizontal: what we predicted. Vertical: what happened. A perfectly calibrated model sits on the dashed line. Each dot is one probability bucket, sized by how many predictions fell in it; the bar is its 95% interval. Wide bars mean not enough games yet — not a wide model.

Put in the same units as the rest of the page: Full draft, players known has a Brier score of 0.2199, and miscalibration accounts for 0.000008 of it — about 0.004%. Nearly all of the error is the coin the game flips and what the model cannot separate, not the probability being wrong. Overall we over-call blue by 0.05%, a direction an average gap cannot tell you because it throws the sign away.

Riot's match history records who played what, and not the order it was picked in. So we model it: every training board is dealt a reveal order, drawn at random.

In solo queue that is a fair model of the real thing. Your pick slot comes from the lobby and has nothing to do with your role — you can be first pick on support — so the order really is close to a random draw, and using one costs little. What a random draw does not reproduce is the pairing: the order shapes the board, because a champion picked last can answer what it sees and one picked first cannot, and we deal the order independently of the board we deal it to. So a late counterpick sometimes lands in an early slot. How much that is worth is the open question, and the only way to answer it is against drafts recorded as they happened — which is what the draft page keeps when a champion select is entered as it locks.

Toward checking it

70 of 750complete sides, counting the smaller of blue and red — the threshold the estimator itself asks for before it will replace the assigned order

Whole drafts recorded

59from 236 champion selects entered; the rest stopped partway or could not be read back unambiguously

Collecting since

Sep 1, 2026the first champion select entered as it happened

Nothing about you is stored to make this: a recorded draft is the champions, their side and the order, with no account attached. The count moves only when somebody enters a champion select while it is happening, so it is a slow number by construction — and the model does not wait on it.

For each draft, where the champion the player actually locked sat in our ranking of the champions that player already plays -- scored the day after, against a history frozen before the game began.

Days scored

282026-09-20 - 2026-10-20

Games

922,493

Players

318,959

Champions per player

7.3

Which model ranked these days

Every table here pools all of these days, and they were not all ranked by the same model. A day counts for the model that ranked it.

8 of the 28 days (2026-09-29 to 2026-10-06, 281,633 picks) were ranked by a model the site was no longer serving. They say how that model ranked, not how today's does.

Scored after their own night: 2026-10-07, 2026-10-08, 2026-10-09, 2026-10-10, 2026-10-11. Each was ranked from the player history kept on that night, so no pick met a history that already held it.

Was the one we put first the one that won?

Each row is one spot in our ranking, not a running total: #2 is only the picks we ranked second, compared with the picks we ranked below them. Each row also counts only players who had more champions than that spot: in a five-champion pool, “we ranked it 6th to 10th” cannot happen and the row would just restate the average. The comparison is the rest of that same pool rather than 50%, because picks inside a pool already win about 52.8% of the time — that number is familiarity, measured, and it is what a recommendation has to beat.

The same picks, by where in the pool they sat

Top band minus bottom band: 5.0 percentage points. Read the intervals beside each row before treating that as a difference -- they are computed per game, not per seat.

What about picks we did not suggest?

Not in our top 10 minus our top 10: -3.7 percentage points.

Measured on 27 days (2026-09-23 to 2026-10-20): the days that record picks outside the pool.

Nobody is assigned a first-time champion at random. It is autofill, a smurf on a new account, a one-trick widening out, a counterpick someone felt forced into. This says what happened, not what anyone should do — and it is measured on the same drafts, the same day, against the same frozen history as the tables above.

Our #1 against their safe pick

For every pick we also wrote down the champion the player would choose with no help at all: the one they have played most in that role, the one they have been playing lately, and their highest-mastery one. Where that was also our #1, following us and playing it safe are the same thing. The comparison is on the drafts where the two differed: did the players who took our #1 win more than the ones who took their safe pick?

27 days (2026-09-23 to 2026-10-20), 921,907 picks. An interval that crosses zero means the two cannot yet be told apart.

Nobody is assigned our pick or their usual one at random: a player who leaves their main may be countering, filling or trying something new. This compares what happened, not what would have happened.

Did the chance we predicted show up?

For picks in their pool we also kept our predicted win chance for the champion they locked and for our #1. Nobody plays both, so the check is this: picks we predicted further below our #1 should win less, by about what the predicted column says. If the “won” column stays flat while the predicted one falls, the gaps we predict are noise.

Our predicted chance for the pick locked, scored on 741,072 picks: AUC 0.5617 · log loss 0.6870 · Brier 0.2469 · accuracy 0.5428 · ECE 0.0096.

Players choose their own picks, and being comfortable on a champion predicts winning on its own. This shows whether picks we ranked highly won more often. It does not show that following the recommendation causes wins.

Ranked only against champions that player already plays. Ranking against all champions would be easier and less honest: the one they picked is usually the only candidate we hold real history for, so it would win on familiarity rather than on advice. A percentile makes a 6-champion pool and a 40-champion pool comparable. Most players keep only a handful of champions per role, so the percentile takes a handful of distinct values and the bands are uneven by arithmetic rather than by chance -- that is the pool restriction working, and smoothing it would hide the measurement.

Predictions this API returned before the games were played, scored against the results. Nobody chose these games for us.

Counting since 2026-09-29. Earlier predictions are kept in the logs but not counted here.

Drafts with a player named

86one champion select counts once, however many times it asked

Ranked solo games scored

33the only games the figures below are computed on

Other queues

1real games in Normal Draft, Flex or Clash: played, not scored here

Remakes

0

Never became that game

50a dodge, or the player's next game was a different draft

We could not look

0Riot refused the lookup until the deadline: ours to fix

Still waiting

2less than a day old, no result yet

Custom and bot drafts left out

22answered for the desktop companion, never measured

Predictions with no player named

5,459usage: drafts tried on the site; no result can ever exist

Mode: Full draft, players known

19Ranked games it answered · 22 drafts that asked it

Mode: Mid-draft, solo queue

33Ranked games it answered · 85 drafts that asked it

What this is a measurement of

These predictions were not all made by the same model. Every build we have shipped since the log opened is in here, and every patch that has been live since. This is which ones, and how much of the sample each is.

Scored games span

Sep 30 - Oct 20, 2026serving since Sep 1, 2026, 43 days of log

Builds behind it

142026-10-03 (7), 2026-10-19 (5), 2026-10-15 (4), 2026-10-13 (3), 2026-10-09 (2)

Patches played on

21620 (21), 1619 (12)

Answering now

2026-10-20not one of the builds measured above — every scored game was answered by a model since replaced, which is what a daily refit does and not a fault

AUC

0.6287± 0.1925

Log loss

0.65120.693 is a coin flip; lower is better

Brier

0.23190.25 is a coin flip; lower is better

Accuracy

51.5%games called correctly at 50%

Calibration error

0.1905mean gap between predicted and observed, 15 bins

Blue win rate here

48.5%~50% in the population; far off means an odd sample