How to read the numbers

The win-chance model

A model trained on ranked games reads the state of a game every minute (gold, levels, items, objectives, kills) and gives each team its chance to win. The model never scores a game it was trained on. We measure a choice by how much win chance moves after it. Because that movement is far less random than a win or a loss, it takes fewer games to show a real effect.

All differences are in points of that chance and written with %. A difference of +1% is about one more win per 100 games.

How the main number is made

Every verdict reads this number. It compares each option with the most picked one across every game, but only in games that looked the same when the choice was made. For runes, spells and starters, those games had the same win chance from the draft. For an item, they had the same win chance and the same minute of purchase. They were also played on the same patch, by players as experienced on the champion and as often its mains. Take a keystone at −2.7%. In games like those it was taken in, it won about 2.7 fewer games per 100 than the most picked keystone.

The games are matched two ways at once. The first predicts how each game would have gone with the most picked option. The second reweights the games with the most picked option to look like the games with this one. The answer is right if either way is (a doubly robust estimate). Every model is checked on games it wasn’t fitted on. The margin counts each player’s games together.

Runes, spells and starters get one more step. Plenty of games are won or lost in another lane, whatever this player picked. So these games are also compared after accounting for how the other lanes were doing. That means how far ahead or behind in gold each other lane was at 14 minutes. At 10 minutes it also means the jungle’s gold and everyone else’s experience and kills. Luck elsewhere on the map then counts for less. The margins on these choices get about a tenth narrower.

A player’s own lane never counts, because their choices change it. A bot laner and their support share a lane. Neither counts the other. Junglers and supports roam. For them we keep only what their choices don’t change. Every rebuild also checks each champion on its own. If its choices change how another lane goes, that lane is left out of its numbers. Items skip this step. The win chance when an item is bought already includes the whole map.

It can’t see a player’s skill beyond their experience on the champion. An option that better players favor can still look a little better than it is. So the number shows what goes with winning in similar games. It can’t prove that the option causes the win. Where enough players tried both options, we also compare each of those players with themselves as a check.

The baseline

Every option is compared with the most picked one. The most picked option can’t be compared with itself, and its tag reads BaselineBaselineThe most picked option. The others are compared with it. Its own number compares it with the next most picked.. Its row shows its closest rival’s comparison, read the other way. Say the second most picked keystone is −2.7% against the most picked. Then the most picked one’s row reads +2.7% against it.

Why numbers from few games are pulled in

The site compares thousands of options at once. Even if none of them mattered, about 1 in 20 would look clearly better or worse by luck alone. The ones that look best are mostly the luckiest. So we first pull every comparison toward a safer guess. How far depends on how noisy the number is, set against how much options really differ beyond luck. A number from few games keeps little of its size. A raw +7% from 300 games can read under +1%. With thousands of games, a number barely moves.

The safer guess is what the same option does against the same most picked option on the other champions. Say Teleport does a little worse than Ignite on most champions whose players usually take Ignite. Then a champion like that with few games starts from “a little worse” and moves from there with its own games. Champions with more games count for more. The guess leans on them only as far as the same comparison really agrees from one champion to the next. The champion’s own games never count toward its guess. When no other champion has the comparison, the guess is the average comparison. It sits a little below zero, because the most picked option is usually a good choice.

When a number’s error is more than three times the usual spread between options, the page says “Not enough games” instead. This is empirical Bayes shrinkage. Baseball analysts use the same idea to regress batting averages toward the mean. We also tried a separate spread for each kind of option, and a pull that lets big differences through. Checked on half the games, neither predicted the other half better.

The margin and the verdict

Every number comes with its margin of error, which is the 95% range. The real difference is very likely inside it. On the bars, the dot is the number and the band is the margin. The black line is zero, meaning no difference. The verdict reads the margin like this:

Better and Worse also need a difference of half a point or more. A sure but tiny difference is Even. Even means the 90% range sits within a point either way. The card of what wins most often uses the same test to show two items as “or”. Colors follow the verdict: blue for better, red for worse, gray otherwise.

Not enough games

Some options have too few games, or a number too noisy to say anything. We hide those numbers, and the verdict reads “Not enough games”. We never put forward an option in under 5% of games, however good it looks. A rare option with a big effect mostly comes from a handful of players or from luck. With a dozen options per slot, some clear a 95% bar by chance.

WPA (“vs expected”)

WPA stands for win probability added. For every game with an option, it compares how the game ended with the win chance the game had when the option was chosen (the draft for runes and spells, the minute of purchase for items). It then averages over those games. +1.7% means games with it were won about 1.7 more times per 100 than the game state predicted.

It asks whether the option does better than what players usually do from there. The comparison is with the usual mix of choices in that spot. Because popular options make up most of that mix, WPA understates them. Two options’ WPAs can’t be subtracted to compare them. WPA counts every player, which mixes what the option does with who picks it. An option that strong players favor looks good whatever it does. That’s why WPA sits next to the main number and never decides a verdict.

Next to WPA is the win chance an option is usually taken at. An item bought at 56% is mostly bought in games that are already being won. That makes its plain win rate look better than it is. WPA starts from that 56% to correct for it.

Lane, mid and late game

Differences are split by when the win chance moved. Runes, spells and matchups use three phases: lane (0–14 min), mid game (14–25) and late game (25 to the end). The three add up to the whole game. Items count only from the moment they’re finished. Their numbers cover the first 10 minutes after the purchase and the rest of the game. What happened before the purchase explains why the item was bought. It doesn’t count toward the item.

How we choose what wins most often

For each decision, we pick an option over the most picked one only if it’s in at least 5% of games and reads Better after the pull. We choose it on one half of the games (split by match), and it must still be ahead on the other half. The best-looking of many options is partly the luckiest, and that luck is unlikely to hold on the other half. Otherwise we pick the most picked option. When a second option reads Even with it, we show both. “What wins most often” is these picks put together. It tells you what the games say. You decide what to run.

Each slot is picked on its own. A keystone’s players may usually buy a different kind of first two items than the ones shown. When they do, the card says so.

Builds with each keystone

Runes are chosen before the game and items during it. A keystone often leads to one kind of build. We name the kind from the first two legendary items’ stats. A bruiser pair gives health, armor or magic resist on both items, with damage on at least one. A burst pair gives ability power and nothing that keeps you alive. A mixed pair has one of each. Tank, crit, lethality, damage and support pairs follow the same rule. The rule was fixed from the items alone, before we looked at any result. Each kind is named on the page by its most common pair, such as Hextech Rocketbelt into Riftmaker.

The shares are plain counts. A setup shows when it has at least 200 games and 5% of the champion’s games. Games that end with fewer than two legendary items have no build and aren’t compared. Many champions have no setup that clears the bar yet.

When two keystones share a kind of build, we compare them on those games only, the same way as every other comparison. The numbers are pulled toward the average comparison. They don’t move any other number on the site. Players pick the items after the rune. The rune and how the lane went can both change which items they buy. These comparisons describe the games. They don’t prove what the rune does. If a keystone looks clearly better overall but the gap goes away within one build, the card Goes with the build says so.

What the data says

We run the same seven checks on every champion. Each one is a plain threshold on the margins.

• Common but costly marks an option taken in at least 10% of games that does clearly worse than the most picked option.
• Too good to be true flags a rare option whose big number mostly comes from one player.
• Theory against real games appears when the simulation puts two items 2.5% or more apart and real games can’t tell them apart.
• Depends on your path shows an item whose answer turns around once you look at what was bought before it.
• Goes with the build flags a keystone clearly ahead or behind overall whose gap mostly goes away among games with the same kind of build.
• Pays off in one phase catches options whose edge sits in one phase. Some win lane and lose that edge later, and some pay off late.
• Turns after lane covers opponents who lose lane but outscale after minute 14, or the other way round.

Damage simulation against real games

For items, a damage simulation scores each candidate at a set minute and level (an all-in, a trade, a teamfight, a wave). The score is in percent against the most picked item. The simulation knows item stats but not how games play out. Next to it is what the same items did in real games from the purchase on. When the two disagree, we go with real games.

Everyone or mains

Pages with builds can read everyone’s games or only the mains’ games. A main is a player whose #1 mastery champion is this one. We know who mains what from the mastery of about 717,000 ranked players from Platinum to Challenger. That covers about half the players in our games. The others count only in everyone’s numbers. Mastery is a career total, but the games are the same as everywhere else.

Mains pick differently. On some champions, mains take a different spell, starter or first item from everyone else. Until a champion has 1,500 games by mains, its mains view shows what they pick next to everyone’s verdicts. Before that, mains alone are too few to judge. From 1,500 games on, every number is the mains’ own and compares mains with other mains.

Tier list

Champions are ranked on the win chance they add over their role’s average champion once we take out who plays them. A model on every game with all ten players known accounts for each player’s mastery, rank and games on the champion, and for the rest of the draft. So a champion doesn’t look stronger because one-tricks play it, or weaker because first-timers do. Tiers: S from +2%, A from +0.75%, B from −0.75%, C from −2%, D below. The “% likely” next to a tier is how likely the champion belongs in it. Under 50% means it sits near a tier border.

Accounting for the players is why the tier list isn’t in win-rate order. Win rate lies lists the champions this moves the most.

Matchups and real counters

A matchup’s numbers are compared with the champion’s average matchup. Zero means a typical lane for the champion. It doesn’t mean the lane is even. +3% in lane means the champion wins about three more games per 100 from the lane against this opponent than against its average one. Because the model’s expectation already includes the rest of the team and the draft, they don’t count for or against the lane opponent. Some champions don’t have full game timelines yet. Their pages show the win rate against each opponent instead, compared with their usual win rate.

The Counter goes further. It’s the number behind the Counter column in matchup tables and behind the Real counters list. A plain matchup win rate mixes the counter with how strong each champion is anyway. A strong champion “beats” most lanes. The tier list’s model already knows both champions’ strength, who played them and the rest of the draft. It says what each game should have given. The counter is how much better or worse the lane’s games went than that. Lanes with few games are pulled toward zero. Real counters are worth about a point of win chance on average.

The Early game tab

The tab adds up one effect per champion and role for the 2v2 from the draft. It takes your mid and jungler’s effects minus theirs. A champion left as “any” counts as zero. We tried adding effects for pairs and matchups too. On games the model hadn’t seen, they predicted no better than the plain sum. That means each champion adds about the same whoever it’s with. The draft moves these numbers far less than they vary from game to game. Most of the early game still comes down to how it’s played.

The Explore tab

Each champion’s Explore tab recomputes everything on your own slice of games. You can filter by patch, rank, region, who plays it (from their mastery), how the game stood at 14 minutes, game length, your side, the enemy team’s damage and tanks, your lane opponent, an ally or an enemy. Your browser runs the same comparisons as everywhere else on the games that match, with the same pull. A thin slice shows “Not enough games” instead of a lucky number. Every filter you add cuts games and widens the margins. Filtering on how a game stood at 14 minutes keeps games that were already being won or lost. Their results then say more about how the game stood at 14 than about the draft.

The Insights lists

The Insights pages rank what the rest of the site measures across every champion. They come in three groups: which champion (win rate lies, easy to pick up, high and low elo, comebacks and throws), which lane and draft (real counters, lane bullies and scalers, safe blind picks, best duos) and which build (items that ride the lead, popular traps, underused builds and more). Each list uses the same numbers as above, with their margins and how sure they are. A note under each list says how it’s computed. Every list can be filtered by role and champion. The lists made from slices of the games can also be filtered by when the games were played and by rank. When nothing is clear yet, the list’s note says so.

Checks that the numbers hold up

We can’t see the true effect of an item. We can check that the numbers behave as true effects would. These checks run on 283 champion and role pages each time the numbers are made. We publish the results as they come out.

The data

The site uses 50,990 ranked solo queue games over patches 16.19–16.20, mostly Emerald to Master. Win rates, pick rates and matchups come from all of them. Runes, items and spells need each game’s full timeline. We have that timeline for fewer games, and 190 champion and role pages have enough of them. Builds, runes, the tier list and matchups use only these patches, because the game was different on older ones. The Early tab uses a wider window to get more games. It says so on the page.