These items will (not) lose you the game

Russian version: https://cyber.sports.ru/tribuna/blogs/spectral/2867372.html

Video version (russian): https://www.youtube.com/watch?v=IVeWQFQjwj4

Christmas is just around the corner, and it’s time to choose your presents. But what are presents, really? They are just items, and to choose the best items you need items stats.

But here’s the problem. Items stats are only available on Dotabuff and Stratz. And Dotabuff only counts endgame inventory, while Stratz doesn’t have any kind of summary, only time sliders for a specific hero.

So it became clear to me: if you want to get something done, you should do it yourself. So I’ve made some items stats for heroes.

This text is based on data of the last 40 days of the patch 7.27d (well, I started writing this text a week before the patch ended), but most of the information should be relevant even after the patch.

You can read more about these features in the December Update patchnote. And here you can see items stats for the last week.

Technical part

To generate items stats for all the newly fetched matches, a bit more than a month ago I started recording information about every purchase event.

This kind of data works better to generate some kind of stats. On Dotabuff, for example, items stats is calculated based on endgame inventory, so the resulting data is a bit skewed, since a lot of core items rarely get it to the late game.

However, I’m removing these three types of items from my data:

  • consumables (support items like wards, dust, smoke are also consumables)
  • parts (items that aren’t purchased by themselves on purpose, only as a part of some bigger item)
  • early game items (bought after 12 minute mark)

I’m also removing purchases made in the last two minutes of the game, so weird meme purchases won’t affect the data. This kind of items won’t bring any useful information anyway.

Because of missing consumables, I’m not recording starting items (yet). And I also don’t have neutral items data, since technically they aren’t purchased.

And what kind stats do I have here? Ohoho…

There are a lot of interesting things. But the main ones are item stats (I should’ve named it better, but it’s items stats inside of items stats) and item build progression.

The first is about items’ winrates (and heroes with these items), as well as average, early and late timings, purchase rates and how it changes over time.

All of this data is available as a big cool table, but there are also even cooler boxplots for it.

It’s a boxplot. The words are “minimal”, “early”, “winrate”, “late” and “maximum” respectively, from left to right.

It’s a chart that shows a distribution of values — in this case it’s item purchase timings — so you can clearly see, what timings are early and late (on a time scale below) and what items are bought earlier than others on average. And aside from that you can see how winrate changes over time (it’s this light line in the box).

It’s even more interesting with winrates. Along with the average win rate of an item there are also:

  • winrate shift compared to games when the item wasn’t bought (positive is better)
  • winrate shift with early timing
  • winrate shift with late timing
  • winrate gradient per minute

Gradient represents how winrate changes with every minute of purchase delay . If it’s deeply negative, then the item and it’s purchase time are essential for the hero.

What items will lose you the game?

And what items are going to reduce your chances to win, compared to game when they weren’t bought?

The biggest winrate reduction can be noticed with Diffusal Blade, Divine Rapier, Blade Mail, Maelstrom and Armlet. Most of these items, however, have relatively high winrates with an early timing.

Items and their winrate deviations: late timing (first value) and early timing (second value)

In the top of the worst items list there are also Hand of Midas, Rod of Atos and Bottle. The latter doesn’t even get to positive winrate with an early purchase timing.

If we’re going to filter out items with relatively low number of matches, it gets even more interesting.

Items with the best effect on winrate

In the top of the best items there are Satanic, Assault Cuirass, Butterfly, Abyssal Blade and Shiva’s Guard. Not that far away there are also Eye of Skadi (+10%), Scythe of Vyse (+9.68%), Daedalus (+9.21%) and Solar Crest (+9.02%).

Items with the worst effect on winrate

Among the worst there are Maelstrom, Blade Mail, Hand of Midas, Mask of Madness. All kinds of boots and Urn are close, too.

The item, that reduces winrate the most when purchased, based on the last month of 7.27d worth of data, is Divine Rapier. Usually Rapier is bought in the late game to “strengthen the victory” or as a “last resort”.

It’s also worth to look into item choices for specific heroes. I’d like to point out Hand of Midas for Ogre Magi: in the stats for this patch it’s straight up one of the worst items across all the hero-item pairs. On average this purchase reduces chance to win by 7.64%, and early Hand of Midas reduces it by 9.1%.

It used to have even worse stats.

Across the worst item choices there also are Maelstrom for Weaver (-20.8%), Ghost Scepter for Leshrac (-9.4%), Diffusal Blade for Slark (-8.4%)… I also found it curious to see here Phase Boots for Spectre (-6.6%) and Orchid Malevolance for Clinkz (-6.6%).

Worst items for heroes

Looking at the items with the highest positive effect on winrate, you may notice Aeon Disk for various support heroes, Assault Cuirass for offlaners, Shiva’s Guard and Octarine Core for casters, and Satanic, Butterfly and Eye of Skadi for carries.

Best items for heroes

Medium items and their effect on winrate

I’m dividing items into three big categories:

  • early (cheap, early game items)
  • medium (1000–2500 gold, midgame)
  • major (“big” items, expensive)

Among the medium items it’s not rare to see that an item has a bad “native” winrate, but also has a significant winrate increase with an early purchase timing.

It may mean that these items should be either bought early or not bought at all. On the other hand, it may also be a sign that something went wrong in the game.

Exceptions are items that are popular among supports. Their winrates are usually somewhere around 50% and aren’t really timing dependant.

Another kind of items that is standing out is upgradable items: Vanguard (Abyssal Blade), Mekansm (Guardian Greaves), Medallion of Courage (Solar Crest), Sange (Heaven’s Halberd), Crystalys (Daedalus), Shadow Blade (Silver Edge). Their numbers are affected by the fact that they are going to be upgraded later on.

But what’s even more interesting — exceptions from all these rules. Medium items without upgrades, but with high winrates are Holy Locket (66.12%) and Vladmir’s Offering (61.85%).

And before we get to Holy Locket and why is it here, we should discuss item timings first.

Timings and anti timings

Timings are very peculiar. Time of purchase may affect the game’s outcome, but can also be a sign of how the game goes (and it’s what results in the same situation with medium items).

“Average timings” by themselves are pretty meaningless, so you need some additional variables to create a representation about purchase time’s distribution.

Main ones are 1st quantile (Q1), 3rd quantile (Q2) and median. Roughly these are early purchase timing, late purchase timing and “standart” timing respectively. There are also minimal and maximum values that let you know how far the borders go. All this was briefly explained before, at the boxplot part.

And here’s where winrates come in: early winrate is based on early purchase timing and earlier, late winrate is based on late purchase timing and later.

Based on this data you may draw conclusions about when the item is bought, who buys it and how it affects the game’s flow.

For instance, Holy Locket has purchase time range of 23–34 minutes, and median purchase time is at 28 minutes. Considering average game to last 35 minutes and the item’s high winrate, it would be easy to think that it’s just purchased as a meme while winning, but… both early and late winrate shifts are negative. It already means that it’s a wrong conclusion. And once you look at the heroes who buy Holy Locket, you understand it all.

At the top there are heroes with very strong healing: Io (14% purchase rate, 67.49% -0.15%/min winrate), Phoenix (1.69% purchase rate, 61.54% + 2.32%/min winrate), Undying (7.22% purchase rate, 60.23% +0.73%/min winrate) and Omniknight (4.59% purchase rate, 67.74% +0.12%/min winrate).

Locket becomes stronger with time only for one hero: Phoenix. However, it’s not the core item for the bird, and the high winrate shift may be related to it’s percentage based healing.

Opposite example of items that are bought «to strengthen» or «for memes» are Kaya & Yasha, (+3.7%..-1.8%), Sange & Kaya (1.9%..-7.9%) and Dagon 5 (+5.8%..-8%). These items also have very low purchase rates, which lowers their ranks.

The most useful thing in understanging item timings is winrate gradient. By just looking at its values you can understand, which items depend on purchase time and which don’t: if gradient is highly below zero, it means chance to win drops significantly as the game goes.

Items with highest negative gradients across all heroes; median timing and gradient

If you take a look at the items with the highest negative gradient across all heroes, you may already notice some interesting items: Battle Fury, Manta Style, Bottle, Wraith Band, Abyssal Blade and Mask of Madness.

On the other hand, looking at the highest negative gradients across item purchases for heroes will ppoint out at Drum of Endurance and Necronomicon for Lycan, Diffusal Blade for Naga Siren, Manta Style for Anti-Mage and Terrorblade, Radiance for Spectre, Hand of Midas for Arc Warden, Armlet for Huskar, Eul’s Scepter for Death Prophet and Battle Fury for Anti-Mage, Troll Warlord and Phantom Assassin.

If gradient is fluctuating around zero or is positive, then an item’s purchase time is likely to be unimportant. A good example of this may be Blink Dagger for Lion: the item’s winrate is 52.97%, and its gradient is 1.24% per minute.

Progression tree

And in the end I’d like to tell you about another cool thing I’ve made.

I wanted to make something about item combos. But overall item combos isn’t that useful and is kind of boring. So I decided to take purchase order into account and collect data about pairs of items that were bought one after the other.

Based on this you may already be able to draw a cool tree of a build’s progression. On this tree every item is a node, and every pair is represented by an arrow between nodes. Direction of the arrow shows the purchase order, its thickness represents its purchase rate, and color represents winrate.

An item’s vertical position is based on its median purchase timing, with every level representing 6 minutes of the game.

Ah, and yes, it’s all available in interactive mode on my website. You can select items to highlight their connections, hover over items and edges to see detailed data and move items around.

By just looking at this kind of a tree you can already see what people tend to build on a hero and in what order.

There are some heroes with a rather obvious build, like Anti-Mage or Spectre. You can clearly see the main thread with the main item build. For Anti-Mage it’s, for instance, Ring of Health, Power Treads, Battle Fury, Manta, Abyssal Blade, Eye of Skadi. For Spectre it’s Magic Wand, Blade Mail, Power Treads, Manta Style (maybe Radiance or Diffusal Blade before or after), Eye of Skadi and Abyssal Blade.

There are some more versatile heroes, who still have a clearly visible primary build, like Pugna. Main items are still noticable: Null Talisman (more than one), Arcane Boots, Aether Lens. But then there are divergences: you can see Rod of Atos, Eul’s Scepter, Glimmer Cape, Blink Dagger, Dagon.

The most interesting heroes to look at are the ones who have more than one distinct playstyle, like Lina or Vengeful Spirit. You can still clearly see a “primary” playstyle, but there’s also another one “secondary” playstyle visible in the background, with a less thick line. Usually it’s division between “support” and “core”.

But how do you think, what hero has the most versatile progression tree? Surprisingly, it’s Viper. His item build is the most flexible in terms of both number of items and number of connections between them.

Closing words

Aaand that’s about it! It’s my small christmas gift for y’all. And, of course, the next material about this topic will most likely be based around the new patch. However, it will happen only in the next year.

Don’t forget to subscribe to me on Twitter (or VK/Telegram), join my Discord and share this text with others.

Merry Christmas and have nice games in 2021!

Writing code and stories. https://spectral.gg/

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store