Why League of Legends Item Win Rates Can Mislead Your Build

A raw item win rate answers one broad question: in games where this item appeared, how often did the buyer win? That is not the same as asking whether the item caused the win.

League of Legends items are bought after the game already has a shape. A player may have more gold, better map control, an easier lane, or a team composition that makes the item easier to use. The item inherits that context.

The cleanest example is a snowball item. If an ADC finishes a third damage item early, that item will often show a strong win rate. But part of that number comes from the fact that the ADC was already rich enough to finish it early.

Defensive and situational items can have the opposite problem. A magic-resist item bought against a fed mage may be the correct buy, but it can still carry a weak win rate because players buy it in games that are already hard.

• Ahead/behind bias: players who are already winning complete stronger and more expensive items earlier.
• Survivorship bias: late-game items only appear when the game lasts long enough for them to be completed.
• Emergency-buy bias: defensive items are often bought after the threat has already become a problem.
• Champion and patch bias: an item can look better because its main users are strong on the current patch.

Baron Buff treats an item completion as a timestamp. Instead of stopping at the final result, we ask whether the relevant gold position improved after the item entered the game.

For laning purchases, we look locally: did the buyer gain ground against the lane opponent after the purchase window? If the item is meant to change the lane, that is where the first signal should show up.

For mid and late game purchases, the signal shifts toward team gold. At that point, items are often about fights, objectives, and team needs, so team-level movement is more relevant than a lane-only view.

A purchase is treated as exposure metadata. It tells us what the player bought in that game state, but the purchase itself is not automatically labeled as good.

Positive relevance comes from positive gold-difference movement after the purchase. Flat or negative movement does not create a positive label. Groups with no positive signal are excluded from ranker training because they do not tell the model which valid item was better.

The model also does not get to recommend anything it wants. Before ranking, deterministic rules filter for item legality, slot pressure, patch freshness, buyer fit, and team coverage. The model only orders candidates that already pass those checks.

Post-purchase gold delta does not establish causality. League of Legends has too many moving parts for one item event to explain the next fight, objective or lane trade on its own.

It is still a more useful coaching signal than raw win rate because it starts at the moment the player actually made the item decision. That makes it less about who was already winning and more about what happened after the buy.