Bureau of Rift Statistics · Report No. 2 · Method

How the census was counted

Report No. 2 asks whether players move further than a champion does when Riot changes it, and whether the pros follow the notes at all. The analysis plan was written and locked on 21 Oct 2026, before any match was pulled. It changed once, on 25 Oct, before any result was opened (below). Report No. 1’s method follows further down.

The sample

202,725 ranked games of 15 minutes or more, NA ranked Solo/Duo (NA1), played 9 September to 25 October 2026, patches 26.18 to 26.21 (a game’s patch comes from its game version, never its date). 18,523 players were drawn at random from Riot’s ranked ladder in three rank groups (Iron to Silver, Gold to Emerald, Diamond and up), in a main wave and a top-up wave. 2,027,250 champion appearances.

Each game is weighted so the sample adds up the way the ladder plays: its rank group’s population over the players drawn in that group, adjusted for the share of the main wave’s games that were kept. On the days both censuses cover (29 Sep to 15 Oct), Report 2’s pick rates sit within Report 1’s margin of error for 158 of 173 champions, and the two lists correlate at 0.997.

The comparison

For each champion Riot changed in 26.19 or 26.20, the two weeks after the patch (days 0 to 13) are set against the two weeks before, and the same change for champions Riot left alone, in the same role and rank group, is taken away. That is an event study, the way an economist reads a company’s news against the market. It is done twice: for how often players picked the champion (on a log scale, reported as a percentage) and for how often it won (in points). Champions need 1,000 games on each side to be reported on their own; cells with fewer than 30 appearances are left out of the averages.

Every range is a 95% range from 2,000 reruns that resample the drawn players within their rank group and wave. In simulation, the win-rate ranges caught the true value 19 times in 20, the pick-rate ranges about 9 times in 10.

Riot’s changes in 26.19 to 26.21 were coded on 21 Oct, before the data, as buff, nerf, adjust or excluded, by Riot’s stated intent, with the audience the note names (pro, elite, or everyone). Skins come from the patch notes.

Checks

• A second method. A model of every game with champion and day effects puts the 26.20 nerfs’ win change at −1.3 points (−2.6 to +0.1), next to the main estimate’s −1.6.
• Who left. Players were sorted by mastery points on each champion (top third: experienced). Mastery was read when the data was pulled, not at each game.
• The worked example. Yunara’s plain rates (one game in eight, then one in twelve) are a crude illustration; the tested estimate is the one above.

The one change before results

The plan’s sharpest test was a “belief slope”: how far picks move for each point of win rate. A simulation of this census showed the slope can’t be measured at this size (it reads near zero whatever the truth, a problem epidemiologists call regression dilution). The test was replaced, before any result was opened, with the side-by-side comparisons in the report. The swap and the simulation are in the analysis plan’s amendment 6.

The stage

Every Worlds from 2019 to 2026 was read on Leaguepedia’s Picks and Bans pages in the browser: 833 games. Presence is the share of games in which a champion was picked or banned. Champion changes in each Worlds patch and the two before it, 414 in all, were coded by Riot’s stated intent. 2026 runs through 26 October.

Matchup spread: K’Sante’s win rate against each lane opponent with 30 or more games, pulled toward his average by a random-effects fit (DerSimonian and Laird), with a Q-profile range. He had under 2,000 games a patch, so the planned safe-blind-pick test was not run; the figures describe two patches.

The figures

Each figure borrows a method from a field that uses it: an icon array (Spiegelhalter), funnel plots (Spiegelhalter et al.; Dover and Schopflocher), an interrupted time series (Lu et al., BMJ 2014), small multiples (Tufte), Stein’s two-line plot (Efron and Morris) and a control chart (Dover and Schopflocher). Each figure names its method on the page.

Where the answers stop

• All five 26.20 nerfs were played on the Worlds stage in the same two weeks solo queue drifted away from them, so reading the notes and watching the pros can’t be separated.
• Cassiopeia had too few games (920 before, 789 after) to be measured on her own.
• Mid-patch updates at past Worlds are not coded.
• Three checks were added after the plan was locked and are labelled so in the report: whether players came back to the 26.19 nerfs within four weeks, whether experienced players won more after a nerf, and the Kog’Maw control chart.

Bureau of Rift Statistics · Report No. 1 · Method

How the census was counted

Report No. 1 asks one question: who spends the most time on the grey screen, and do escape tools keep champions off it? The analysis plan was written and locked before the data was opened. It was run once, on 15 Oct 2026. This page says what was counted, how, what was checked, and where the answers stop.

The sample

• 30,510 ranked games of 15 minutes or more, from 7,717 accounts, NA ranked Solo/Duo (the NA1 server only), played 29 Sep to 15 Oct 2026, patches 26.19 and 26.20 (Data Dragon 16.19 and 16.20).
• Drawn like a poll. Players were drawn at random from Riot's ranked ladder, about 1,000 per tier from Iron to Diamond and 250 from each of Master, Grandmaster and Challenger. Draws continued until each tier had enough players who had played ranked in the window. In Bronze, about three in four accounts drawn had not. Up to five ranked games were taken from each player's two weeks.
• 305,100 champion appearances. The analysis frame has 314,830 rows because 973 games were drawn by two sampled players and appear once per draw.

Weighting

The sample is lopsided in two ways, and both are weighted back:

• Tiers. Gold has far more players than Challenger, but a similar number was drawn from each. Each tier is weighted by its population over the accounts drawn (20,464 draws, of which 7,750 had a ranked game in the window and were kept).
• Players. Everyone gave up to five games, whether they played five that fortnight or twenty. Each game counts for the games its player actually played, capped at twenty, so one marathon account can't swing a tier.

Estimated players active in the window: about 589,000 of 1,685,002 ladder accounts.

What "grey" means

• Grey: the share of time played spent dead, weighted by game.
• Escape tools: every dash, leap, blink, untargetability, stasis and revive in all 173 kits, from Riot's own ability text. These were checked against the community wiki's lists and then by hand. The first keyword pass found 71% of movement tools and 39% of targetability tools. The wiki cross-check caught this before the data was opened. Readers later asked whether dashes that only go toward an enemy should count at all. Follow-up 1 splits them, with every call and its sources.
• Ranked champions: a champion is judged in a tier only with 200 effective appearances and 30 distinct players. That gives 65 to 82 champions per tier.
• Intervals: 95%, from 2,000 Rao-Wu bootstrap replicates (the reruns behind every "give or take").

The questions, and which were set in advance

Exploratory results were chosen after the main tests had been run. The report says so wherever it uses one.

Checks run before anything was written

Where the answers stop

• K'Sante can be judged only at the top. He is rankable in Master to Challenger alone. "Unkillable" can't be tested where most players live.
• Regions are description, not an effect. Regions differ in class mix. Adjusting for class (E7) shrinks most of the differences to nothing.
• The death-count model misses deathless games. It predicts 1.2% deathless appearances against 2.3% observed, probably the enchanter supports and Yuumi. Read its rate ratio with that limit.
• Not yet run: hazard by game minute, and the longest single grey stretch.

Changes since the data was opened

• The tier-level intervals and the length-adjusted gradient were computed in a separate script. No rule changed.
• The whole run was repeated on NA1 only after 35 games from other Americas servers were found. Every reported figure moved by 0.02 points or less. See Corrections.

The numbers in each figure can be downloaded from the data page. Every number in the report has a row in the claim trace: the figure, the file it was checked against, and its status.