Findings

What 30,000 ranked games say about champion select

Six charts. Each is drawn from the published aggregates or from the held-out test described on the method page, and every chart has its numbers underneath. EUW ranked solo, Silver to Platinum, one patch at a time. On a phone the charts scroll sideways.

When it says 58%, does that side win 58% of the time?

The model was built on one half of the collection and then asked to score every slot of the other half: 15,037 matches it had never seen, 150,370 predictions, using nothing but the ten champions and their roles. Each dot is one band of displayed numbers; its height is how often that side really won; its size is how many predictions fell in the band.

40% 40% 44% 44% 48% 48% 52% 52% 56% 56% 60% 60% perfect calibration says 41%, wins 44% says 59%, wins 58% what PocketPick displayed before the game how often that side actually won

The dots sit on the line. The fitted slope is 97% of perfect calibration. The band is narrow on purpose: 69% of all predictions fall between 47% and 53%, because a draft alone rarely decides a game. Against a coin flip the Brier score improves by only 0.27%, which is the honest size of what champion select can tell you before the game starts. Source: patch 16.14 collection, 29,970 games, split by match id parity.

The numbers behind this chart
DisplayedActual win ratePredictions
41.2%44.1%1,461
43.2%43.5%5,087
45.2%46.8%14,972
47.1%47.8%32,856
49.0%50.3%45,440
50.9%51.9%31,214
52.8%54.5%12,498
54.8%55.4%4,198
56.8%56.8%1,515
58.7%57.6%450

What actually moves the number

The displayed number is three signals added to 50%. For each signal, the held-out predictions were sorted into fifths, from the most negative to the most positive, and the real win rate of each fifth was checked. The distance between the end dots is how much that signal is worth in practice.

46% 48% 50% 52% 54% Lane matchup 53.2% +6.0 pts 47.2% Champion strength in the meta 52.1% +4.8 pts 47.3% Team comp 51.1% +2.7 pts 48.4% actual win rate, from the bottom fifth of the signal to the top fifth

Lane is worth about 6 points, champion strength in the meta about 5, team comp about 3. Comp is real, monotone and the smallest of the three, which is why the app caps it at 4 points and why a comfort pick usually beats a comp pick at these ranks. Same source as chart 1.

The numbers behind this chart
SignalBottom fifth2nd3rd4thTop fifth
Lane matchup47.2%48.6%49.8%51.1%53.2%
Champion strength in the meta47.3%49.7%50.1%50.8%52.1%
Team comp48.4%49.5%50.4%50.6%51.1%

Scaling drafts change which games you lose, not how many

Each champion and role got a scaling index on the training half: its win rate in games of 30 minutes or more minus its win rate in games under 25. On the held-out half, the 6,015 games where one draft scaled clearly harder than the other were sorted by how long they actually ran. The bar is how often the scaling-heavier side won.

36% 40% 44% 48% 52% 56% 60% even 42.0% under 25 min 18% of games 48.4% 25 to 30 27% of games 51.8% 30 to 35 29% of games 55.2% over 35 26% of games game length

Across all lengths the scaling side wins 50.0%: nothing. It loses badly when the game ends early and wins clearly when it runs long. Long games are about three times as common as short ones (mean length 31 minutes), so the two cancel. Drafting to scale does not make the game longer either. Strongest scalers in the collection: Kayle top (+22 points long versus short over 322 games), Aurelion Sol mid (+21); strongest early champions: Mel Adc (-20), Trundle top (-18).

The numbers behind this chart
Game lengthScaling side wonGamesShare of games
under 25 min42.0%1,07618%
25 to 3048.4%1,61627%
30 to 3551.8%1,74329%
over 3555.2%1,58026%

Your record starts to outweigh the meta after about 45 games

The app blends your own win rate on a champion with that champion’s win rate at your rank. The blend treats the meta as worth 45 games of evidence: effective rate = (your wins + 45 × meta rate) ÷ (your games + 45). The curve is the share of the number that comes from you.

0% 25% 50% 75% 100% 0 50 100 150 200 10 games: 18% you 45 games: half and half 150 games: 77% you games you have played on the champion weight of your own record

Going 7 and 3 on a champion reads as 53.6%, not 70%. At 10 games your record carries 18% of the number; at 45 games it is half; at 150 games it is three quarters. The 45 was chosen slightly below the statistical optimum so that a real record eventually wins the argument. This is arithmetic, not a measurement, but the 45 came from checking personal records against outcomes.

The numbers behind this chart
Games on the championYour recordMeta
510%90%
1018%82%
2031%69%
4550%50%
9067%33%
15077%23%
30087%13%

Gold at 15 and matchup win rate barely agree

Each dot is one lane matchup with at least 60 games in the collection (485 matchups), placed by the average gold lead at 15 minutes and how often that side won the game. Every pair is shown from the side that led in gold.

20% 30% 40% 50% 60% 70% 80% 0 +250 +500 +750 +1,000 +1,250 +1,500 even 1 1 Illaoi vs Nasus: +1,500 gold at 15, wins 60% 2 2 Zed vs Veigar: +1,100 gold, wins 50% 3 3 Draven vs Jinx: +1,400 gold, wins 42% average gold difference at 15 minutes in that matchup matchup win rate

Correlation 0.14. Winning lane and winning the game are related, but weakly: Draven leads Jinx by 1,400 gold at 15 and loses the game 58% of the time; Zed leads Veigar by 1,100 and wins exactly half. In the held-out test, matchup win rate predicted the result better than gold at 15 did, and because the two barely overlap the app uses both, at 60/40. Source: the published patch 16.16 aggregates.

The numbers behind this chart

The eight matchups with the largest gold gaps (the full set is in matchv5-metrics.json on the data page):

MatchupWin rateGold at 15Games
Nasus vs Illaoi (Top)39.6%-1,533101
Draven vs Jinx (Adc)42.3%+1,40271
Garen vs Jax (Top)61.0%+1,238123
Veigar vs Zed (Mid)50.0%-1,10398
Garen vs Tryndamere (Top)45.5%-1,09399
Garen vs DrMundo (Top)58.7%+1,02563
Teemo vs Yorick (Top)45.3%-99875
Mordekaiser vs Malphite (Top)56.1%+98782

Champion synergy between strangers: nothing to see

Every pair of champions on the same team, ten per team, about 597,000 pair observations across the collection. For each pair with 20 or more games (5,062 pairs), the residual is its actual win rate minus what its two champions’ own rates predict, shrunk for sample size. A real synergy would sit well to the right of zero.

0 90 180 270 360 -6 -4 -2 0 +2 +4 +6 zero: the pair does exactly what its two champions predict pair residual, percentage points of win rate number of champion pairs

Centred on zero, and the sum of a team’s pair residuals does not predict its held-out games (z = 1.4, not significant). The pairs at the far ends have 20 to 25 games each and are noise. The combinations people swear by: Malphite and Yasuo −5.4 points over 154 games; Leona and Miss Fortune +6.1 over 176; Nautilus and Samira +5.3 over 131. The positive ones are bot-lane duos, and the data cannot tell friends on voice from strangers, so the app scores no synergy at all: it asks whether a pick covers a job the team is missing.

The numbers behind this chart
Residual (points)Pairs
-6.0 to -5.566
-5.5 to -5.079
-5.0 to -4.5114
-4.5 to -4.0149
-4.0 to -3.5159
-3.5 to -3.0189
-3.0 to -2.5232
-2.5 to -2.0233
-2.0 to -1.5286
-1.5 to -1.0299
-1.0 to -0.5297
-0.5 to +0.0321
+0.0 to +0.5288
+0.5 to +1.0269
+1.0 to +1.5275
+1.5 to +2.0269
+2.0 to +2.5246
+2.5 to +3.0202
+3.0 to +3.5207
+3.5 to +4.0150
+4.0 to +4.5134
+4.5 to +5.0116
+5.0 to +5.586
+5.5 to +6.081

What is missing

Charts 5 and 6 can be recomputed from the files on the data page; charts 1 to 3 come from the held-out test described on the method page; chart 4 is the formula itself. If a number does not hold up, hello@pocketpick.gg or the Discord.