How we compute item advice
We estimate what an item changes at the moment you buy it, not at the end of the game
Item win rates are the most misleading number on every stats site, and the reason is not sloppiness. It is that the number answers a different question than the one you asked.
Inside 10 strata of the game state. That estimate runs in the app, which can see your game. This website has the measured item pages and the patch boundary checks behind it.
You buy an item because of the state of the game, and the state of the game decides who wins. So an item win rate counted at the end of a game is measuring the games in which people bought that item, not the item. Finishing your sixth item is a symptom of winning.
The fix is to ask the question at the moment of the decision. Two players in the same position, same gold difference, same side of even, one buys the item and one does not. That comparison is the one that answers what buying it does.
These numbers come from 108,616 ranked games we recorded on EUW, Korea and NA, found by following Diamond players and the people they play with. That is not every rank and it is not every region. Everything on this page is patch 16.14. Counted up to .
Why do other sites get item win rates backwards?
This is not a small bias. The sign flips.
The trap
Simulate it. Let one hidden advantage drive both winning and finishing items, and set every item's real effect to exactly zero. Count win rates at the end of the game and an item that does nothing shows about 92% as a sixth completed item. That is the survivorship in the number, on its own, with the item contributing nothing at all.
Now run it the other way. A defensive item with a real 5 point benefit, bought mostly when a player is behind, shows up about 20 points in the wrong direction. The raw table says the item loses games. The item wins games. Every stats site's item list has this shape in it, and it buries exactly the purchases a losing game needs.
Counting at the end, against measuring at the buy
What other tools do
Count items at the end of the game and sort by that number. "59% of Jinxes finish Berserker's Greaves" tells you who lasted long enough to finish boots.
Fair credit where the market has it: Coachless builds its headline metric on the state at the moment of purchase, which is the right conditional. One other tool states that it corrects for this but has not published the correction, and a correction nobody can check is worse than none, because it puts a clean label on a biased number.
What we do instead
Estimate what buying the item changes at the moment you buy it. Two players in the same position, same gold difference, same side of even, one buys the item and one does not. That comparison is the one that answers what buying it does.
What do you estimate at the moment you buy?
The rules, in full, so you can argue with them.
Every purchase in the store carries the state it happened in. We cut that into 10 strata: whether the player was ahead, crossed with the gold difference at the moment of the buy in buckets at 2000, 500, 500, 1500 gold. Inside a stratum, buyers of the item are compared with a neutral background stream.
The background stream is the long tail: items with fewer than 50 purchase events each. They are bought too rarely to carry a selection story of their own, which is what makes them a usable comparison rather than another biased column.
Two models back each other up. One estimates who buys, one estimates who wins, and the estimate stays honest if either one of them is right. Buying chances are held between 0.1 and 0.9, so a stratum where almost nobody buys cannot swing the whole answer through one player. An arm without enough events does not get a confident number, it gets a wide range that says so.
What comes out is still an association, adjusted for the state we can see. Two players in the same gold bucket can differ in ways match data does not record. We say that here rather than calling the output causal, which is the line the rest of this market steps over.
What happens when an item changes in a patch?
6 items on patch 16.14, each measured 10 days either side of the boundary, so the players stay roughly the same and the patch is what changed.
When Riot changes an item, there is a second way to measure it that does not depend on any adjustment at all: compare the games just before the patch with the games just after. The population barely moves across a boundary like that, so what changed is mostly the patch.
| item | change for buyers | games each side |
|---|---|---|
| Hextech Rocketbelt | +0.38 wins per 100 games (range -0.33 to 1.09) 85,554 games Measured | 28,390 |
| Kraken Slayer | -0.23 wins per 100 games (range -1.13 to 0.66) 52,094 games Measured | 19,062 |
| Phantom Dancer | -0.3 wins per 100 games (range -1.6 to 1.0) 27,492 games Measured | 8,637 |
| Rapid Firecannon | +0.3 wins per 100 games (range -1.0 to 1.6) 24,308 games Measured | 8,050 |
| Navori Flickerblade | +0.7 wins per 100 games (range -0.6 to 2.1) 23,914 games Measured | 7,998 |
| Protoplasm Harness | -1.5 wins per 100 games (range -2.9 to -0.1) 21,570 games Measured | 7,011 |
Read the ranges before the values. On this patch one of the 6 ranges sits clear of zero, and the rest cover it. A range covering zero is not a failed measurement. It is the measurement telling you the change is smaller than the games can resolve.
The boundary check is the half of this that a web page can carry, so it is on this page and on the item page for every item that has one. The adjusted estimate is computed from your game, so it is the app's. Neither gets dropped for disagreeing with the other: when two methods point different ways about the same item, that is the interesting part.
Where the buy answer actually lives
Plainly: not on this website yet.
The buy board reads your champion, your inventory, your gold and the enemy team, and that takes a signed request from a program that can hold a secret. A static web page cannot hold one, so the live buy answer runs in the desktop app and not here.
What is on this site is the measured store behind it. Every item has a page under the item list with who finishes it, when it gets bought and what the current patch changed, each with its games. Those pages count what happened. The buy board in the app is the part that answers what to do about it, in front of you, at the moment of the buy.
How to check us
-
Open https://api.ko.lol/v1/rdd?patch=16.14. Every report carries its range, its total games and its games on each side of the boundary.
-
Open any item page on this site and count what sits next to each number: the games, the range, and the word for where it came from. Then look up the same item on any aggregator and count what sits next to theirs. That is the whole comparison, and you can run it on this website today.
-
Check the rules before you check the output. The bucket edges, the clip band and the background stream are stated above, in full, so the moment of the buy estimate can be argued with before anybody installs anything. The estimate itself is in the app, because it is computed from a game this website cannot see.
-
Watch a boundary report over the first days of a patch. The range narrows as games arrive. If it does not narrow, the games are not arriving, and the report says how many there are.
What we do not claim about items
The estimate is an association adjusted for the state we can see, not a proof of cause. Inside a gold bucket, buyers can still differ from non buyers in ways the match data does not record, and no amount of adjustment fixes a thing that was never recorded.
Component and boot shares are frequencies: what players in your situation bought, counted. They are not claims about what those purchases did.
ko.lol for desktop
Your champion pool, your matchups and your patch notes, on your desktop while you play.
Free. For Windows.