Companion to the aggregate data and the findings charts. Written because people asked what sat behind the findings, which was a fair thing to ask.
It is not a trained classifier. It is an additive model, and every term is in percentage points of win rate, so the total can always be taken apart. The displayed number is three things added together and then clamped:
| You and meta | The champion's win rate in the current meta, blended with the player's own record on it. The blend treats the meta as worth about 45 games of evidence, so below that a personal record barely moves the number. |
|---|---|
| Lane | How far the matchup deviates from even. Roughly 60% matchup win rate, 40% gold difference at 15 minutes. |
| Comp | The team's composition against the enemy's, capped at ±4 points so it can never dominate the total. |
Some parameters were fitted by split-half testing; some caps are hand-set. Which is which is flagged in the app itself. The model is additive rather than a black box on purpose: every number opens into a panel showing what produced it.
| Source | Riot Match-V5, matches plus timelines |
|---|---|
| Queue / region / band | 420 ranked solo, EUW, Silver to Platinum |
| Patch / size | 16.16, 30,000 matches |
| Excluded | Games under 8 minutes, so remakes and early surrenders |
Yes, slightly. Roughly +1 percentage point per standard deviation of comp quality, z=2.20. Split into fifths on held-out games, top fifth against bottom fifth is about 3 points of actual win rate. The same split on lane quality gives about 6 points and on champion strength about 5, so comp is the smallest of the three. The engine caps the whole comp term at ±4 points, which on this evidence is generous rather than conservative.
Three steps, and the first two matter as much as the third.
Limitation, stated plainly. Match-V5 carries no party flag, so premade status is never observed. Bot lane is a proxy for it, and a confounded one: an ADC and a support share a lane and interact more than any other pairing whether or not they are friends. The effect is real and it lives in bot lane. Calling it premade coordination is an inference, and "two people in the same lane interact more" fits the same data equally well.
Which champions count as scaling was measured, not decided. Per champion and role, the scaling index is win rate in games of 30 minutes or more, minus win rate in games under 25, shrunk for sample size and derived on the training half. Kayle top came out +23 points across 352 games, Aurelion Sol mid +24 across 233; Ziggs ADC went the other way at −24 across 497.
On 5,947 held-out games: short games, scaling teams lose at z=−3.18 (−5.2 points per standard deviation). Long games, they win at z=+1.85 (+1.6). Long games are about three times more common, so the two roughly cancel and the net is z=+0.54, not significant. Drafting for late game also does not make the game longer: duration against scaling gives t=0.17, effectively nothing.
So scaling is a real mechanism and a poor win lever. It changes which games you lose.
The blend uses a prior worth K games: effective rate = (wins + K × prior) / (games + K). At games = K the record and the meta carry equal weight. K was calibrated at 45, chosen deliberately slightly below the statistical optimum so that a real record eventually leads. Worked example: 7-3 on a 50% champion reads as 53.6%, not 70%.
Tested leakage-free on 29,674 games by deriving both on one half and predicting the other. Whole-game matchup win rate z=+2.33, gold difference at 15 z=+1.81. Included together, win rate holds at +2.02 and gold at 15 at +1.38. They correlate at only 0.18, so blending beats either alone. The ratified mix is 60% win rate, 40% gold at 15. The measured slope is about 3.4 points of win rate per 1000 gold, so a 300 gold matchup edge is worth roughly one point.
An earlier run on 11,946 matches produced the opposite ranking. The 30,000-match re-validation reversed it. That is the clearest reason not to trust results built on a few thousand games, including the earlier ones here.
The last test is whether the finished number means what it says. Gold-at-15 aggregates and pair residuals were rebuilt from one half of the collection; 15,037 matches from the other half were then replayed and scored with the engine's own display function.
Displayed 58.7% corresponded to an actual 59.18%. The calibration curve is monotone and tight across the whole output range, with a logit slope at about 97% of ideal.
Three things this test is not. It scores the finished lobby, all ten champions locked and all five lanes known, which is the easy version of the problem; picking third with two enemies unrevealed is harder and has not been measured. It scores all ten slots in each match, so 15,037 matches is 150,370 predictions which are not independent, and the honest error bars are wider than that count suggests. And it ran with no player history at all, so it validates the meta, lane and comp terms and says nothing about question 4.
Worth being clear about what calibration is not. Outputs span roughly 41% to 59%, so on most drafts the model says something close to a coin flip. Scored on how often it picks the winner it would be barely above 50%. The claim is that the numbers mean what they say, not that they are powerful — a draft can only explain a few percentage points, which is question 1 restated.
Found a hole? That is the point of publishing this. hello@pocketpick.gg, or the Discord.