How Player Insights Works

Your Habits tab lists up to 24 habits, from CS at 10 to how you queue after a loss, each with an LP number. This article explains where those numbers come from, what they mean, and what they can't tell you.

The short version: we take each player's habits over five patches and measure how far they climbed over the next four (about two months), comparing players who started at the same rank with similar experience and earlier performance. One model learns how much each habit mattered and how that changes from player to player. Your list is ranked by what mattered most for players like you.

The question

The obvious way to find out what wins is to compare high-elo players with everyone else. It doesn't work. High-elo players are different in almost every habit at once: they farm better, die less, build better and play more. Comparing them can't tell you which of those differences gets you there, and which are just what good players look like.

So we ask a narrower question:

Asking about the next four patches matters. A habit measured over the same games as your rank partly just describes how good you already are. A habit measured before the climb it's compared with can't be a side effect of that climb.

The data

• Players: 3.7 million ranked players with at least 20 ranked games in five consecutive patches (16.11 to 16.15) and at least 10 ranked games in the four after it (16.16 to 16.19).
• Habits: measured from each player's ranked games in the first five patches, from match results, builds, timelines and minute-by-minute positions.
• Outcome: the player's rank at the end of the next four patches, on the LP ladder (one division is 100 LP).
• Training and testing: the model trained on 3.2 million of the players. The other 460,000 were held back, and every check below is on them.

The 24 habits

Each habit is measured against the right yardstick. CS at 10 is compared with your role's average, objective presence with players of your role and rank, early damage with players of your role and rank who climbed. The Habits tab shows the exact definition of each one.

We left some habits out on purpose. Raw stats like KDA or damage per minute mostly measure how good you already are, so they can't tell you what to change. A few habits that looked useful gave advice that flipped between kinds of players, or that no one could sensibly act on, and those went too.

What we hold fixed

For every habit, the model compares players who are alike in:

• Rank: their rank at the start and at the end of the five habit patches.
• Experience: games played, days played, games on their champions, and games in the next four patches.
• Earlier performance: kills, deaths, assists, kill participation, vision, CS, damage and gold per minute against their role and rank, and their win rate.

If one of those performance stats measures the habit itself (say, CS per minute for CS at 10), it's left out for that habit, so the comparison doesn't erase the very thing it's measuring.

The model

For each habit we use double machine learning. In plain terms:

1. Predict each player's later rank from everything we hold fixed, and keep only the part the prediction misses.
2. Do the same for the habit: predict it from everything we hold fixed, and keep the part that's left over.
3. Ask how the two leftovers move together. If players with more of the habit than expected also climbed more than expected, the habit went with climbing.

The predictions in steps 1 and 2 are gradient-boosted trees, each scored on players it didn't train on. Step 3 isn't one number for everyone: the effect is allowed to change with the player's other habits, the roles they play, the kinds of champions they play and their tier group. That's what makes the list personal. CS at 10 can matter a lot for one Silver top laner and very little for a Diamond support.

1from econml.dml import LinearDML2from xgboost import XGBRegressor34model = LinearDML(5  model_y=XGBRegressor(n_estimators=100, max_depth=3),6  model_t=XGBRegressor(n_estimators=100, max_depth=3),7  cv=3,8)9# Y: rank after the next four patches. T: the habit, in standard deviations.10# W: what we hold fixed. X: what the effect may depend on.11model.fit(Y=later_rank, T=habit, X=player_profile, W=held_fixed)1213effect_for_each_player = model.effect(player_profile)14

Anchoring the numbers

A model like this is good at telling which players a habit matters more for. It is less reliable about how big the effect is on average. So the size comes from a direct measurement, and the model only spreads it out.

For each habit we measure the average effect directly, across all the players whose games measure it: the extra climb over the next four patches per one-standard-deviation change in the habit, holding the same things fixed. That's the habit's anchor. Your effect is the anchor plus how far the model's effect for you is from its average. Averaged over everyone, each habit's effect equals the measured one exactly. Yours is higher or lower depending on how much the habit mattered for players like you.

What "+N LP" means

• On a card: the LP difference, for players like you, between your level of the habit and the typical player's (the median). A weakness shows how much further players like you climbed at the typical level. A strength shows how much your level is worth over typical.
• In the detail view: the same number with a 95% range, and what moving to the 50th, 75th or 90th percentile went with.
• The unit: LP over the next four patches, about two months. One standard deviation is a realistic change: roughly what separates a typical player from one in the top sixth.

These are averages for players like you, not a promise for any one player. If you change a habit, your own result can differ.

How we checked it

Every check is on the 460,000 players the model never trained on.

• Your top habit holds up: 86% of the time, the habit we'd tell a held-out player to fix first is backed by a separate measurement among players like them (same main role and tier group): those players climbed further when they moved that way, with a 95% interval that excludes zero. On the 360,000 held-out players both models can be scored on, with the same test, it was 44% for the model we used before and 86% for this one.
• The numbers are the right size: habit by habit, the average number we'd show held-out players is 1.00 times what we measure directly on those same players (the median over the 24 habits; 1.00 would be exact). Split further by main role and tier group, the median is 0.79, so for one kind of player the numbers tend to run smaller than what we measure for them.
• The app computes what we trained: for 22,851 held-out players, the Habits tab's own code gives the same numbers as the trained model, to the 15th decimal place.

What it can't tell you

• It's still a comparison. We compare players who look alike on rank, experience and earlier performance. Something we can't measure, like how much someone practices outside ranked or how motivated they are, could still be behind some of the climb. The numbers say what went with climbing among similar players, not that a habit will raise your rank by that much.
• Habits overlap. Players who deal more early damage often also trade more efficiently. Each habit's number partly includes the habits that go with it, so the "Top 3 fixes added up (they overlap)" number on the page overstates what fixing all three at once would add.
• It's about the next four patches. The model learned from patches 16.11 to 16.19. A big patch can change what matters, and we retrain to keep up.
• Sometimes your advice runs the other way. For about one habit in six that we show, the model finds that players like you climbed more doing the opposite of what helps on average: less of it where most players gain from more. We show you that direction, because it's what went with climbing for players like you.