Russian version

There were multiple reasons for this sort of pause, major ones being my work with Tundra Esports for the past couple of years up until the end of TI12, technical challenges for introducing some new features and general lack of time. I want to cover the first two topics sometime later in my blog, and right now I want to talk about one feature I was working on for a while now: Starting Items and Consumables.

While it's not that fancy of a feature (it's not that exciting to see starting items), there were some things about it I wanted to do a slightly different take on. And Consumables became an interesting byproduct of it, which, I am sure, has a lot of potential.

So let's dive into the story of this feature, how to use it, what it reveals and what can be improved.

(this blog post will also be the first to be published to Substack directly, though all the older posts from Medium should also be there)

…initially this blog post was planned for publishing a bit earlier, but I figured it's better to wait until 7.38 releases and the dust is settled after that.

The Why and The How

The Why is rather simple. At first, I wanted to add starting items to my stats-based builds, but couldn't really find anything that would look at starting items in the way that made more sense to me – by looking at the full builds and looking at the items in the context of laning.

As a byproduct of it came the consumables stats section. The core idea of it was to expand on the laning stats and see, how various "good" players on the roles tend to behave in terms of consumables at the different stages of the game (e.g. how much regen a certain player prefers compared to others and how the team generally behaves in terms of laning habits).

And so the work started, and it took a bit too long.

Well, technically I wasn't really working on it for THAT long. The core prototype was created sometime during late spring of 2022 (and late summer for consumables), then presented as one of my work-in-progress features to the team I was testing for during the TI11 run. It was abandoned for some time. Then picked up again during summer 2023, put on hold until post-TI12 to collect new batch of data after fixing some issues with data processing , when I started asking pro players and talents for feedback. In the beginning of 2024 I spent some time polishing this feature and cracking related performance issues, getting ready to release it in the spring, but had to postpone it due to personal matters. And finally picked it up again during the winter break recently.

So while it was in the works for a while, it didn't really change much for the past two years and was available as an Early Preview feature. Even the code for data processing didn't really change that much after the first major version. Because of that, starting items, builds and consumables sections are supported by all Immortal Rank Meta Trends reports since 7.35d and some competitive events before and after that (though data quality may differ from immortal rank report to report).

There were two major problems: (1) legacy system components, which ended up causing slowdowns and increased resources usage while processing data for this feature in particular and (2) data source issues.

The first problem was popping up during the very long patches and was mostly related to how my system was set up about seven (DAMN!) years ago (when the stats hub was a glorified collection of my "Competitive Meta Trends" reddit posts). But there is also another reason: MMR inflation at the Immortal rank (caused by multiple factors, including Double Down tokens), which caused the increase of daily matches at Immortal (after filtering out matchmaking abuses and botmatches) from ~7k to ~20k, and it also significantly increased the dataset size. Thankfully, I was able to tame it down through some clever hacks and regular monitoring throughout the 2024.

The second problem is of different nature. While I use a combination of OpenDota and Valve's APIs for competitive matches, for ranked specifically I'm relying on Stratz as a provider. Their awesome API was a real lifesaver that enabled all my crazy experiments throughout the years. Unfortunately, they encountered several issues, that seemed like targeted attacks and API abuse, which ultimately led to them limiting query complexity down. This also impacted some calls I was making in a way.

One big problem is caused by Immortal Draft: as the players get "assigned" to their slots, they tend to be slightly mixed up, and this affected the quality of data coming from Stratz. I only noticed while preparing the feature for release in late 2023. In simple words, it meant that initial purchases (aka starting items) were recorded to incorrect players, resulting in practically garbage data. And as in 2023 this drafting mode was fairly new and everything worked fine prior to that, I had to quickly find a new way to reconstruct starting purchases.

The approach I chose was to use a snapshot of players' inventories at minute zero. Howevee, as Stratz was facing these targeted attacks, I had to remove some data from the requests – and this affected backpack slots and charges information in the inventory snapshots, meaning there are only up to six items recorded, and multiple purchases of stackable items are still recorded as one.

While this compromise didn't sound as exciting, it was still a decent way out of the situation, and it worked well enough for my purposes. And competitive stats were largely unaffected by this. I'll get back to this topic later, covering the further improvements to the feature, but for now…

Core ideas

So what's that different about it anyway?

Let's take a look at the starting items section first, as it should be fairly familiar and simple.

All the new sections have support for different roles and "total" for all the heroes playing on this role. The approach to calculating the number of matches played on a specific role is slightly different, so the numbers might not match with the main "positions" section, but the difference is usually very minor.

Starting items section lists all the items that were found in the starting builds, including multiple instances of a single item (unless it's a stackable item, in which case there is a limitation, mentioned prior).

Every item has the number of matches when this item was purchased, overall purchase rate, winrate and also lane winrate (helping to see average laning impact). Toggle filters allow to hide any sequential purchases and purchases with the low number of matches.

In most cases for Immortal Meta Trends reports there is a limiter working to filter out low purchase rate options (which are likely to be noise, replay parsing issues, or both).

Another place of interest is the "Starting Builds" section. The core idea of it is to take the actual full builds and compare them, since sometime starting build is not about any individual item, but rather about the idea that ties everything together. As a nice bonus, there is an option to look at the total numbers for a given role, and find out the "average" build.

The same data is also available for individual players (in non-ranked reports) and even whole teams.

Observations

There should have been some sort of observations block in this text, but frankly speaking there isn't much to talk about.

The most peculiar part for me personally is the meta for each role overall.

The main thing is having a confirmation for trends we already observed for quite some time. For safelane it was already rather common to go for a complete Magic Wand and it's still holding strong, being at the top with 17% purchase rate and whopping 56% lane winrate. Complete Wraith Band (49.79%), Gloves (51.13%), Blood Grenade (50.45%) and Orb of Venom (47.54%) are, on the other hand, some of the worst items in this regard, with Orb of Frost (53.45%) being only slightly higher, though still having a formidable lane winrate.

Surprisingly though, Blight Stone is also very high on the list in terms of lane winrate (I wonder, if it's because of Nature's Prophet?)

Another trend for Safelane Core is to get multiple Circlets, with any number of them having at least 54% lane winrate. We already saw the Circlet trend for a while, but it's nice to have a confirmation of sorts.

Midlane trends are slightly different. First of all, it's usually Midlane players who are guilty of spamming purchase events and skewing the data, which makes navigating their stats a bit… rough.

Unlike Safelane, complete Magic Wand is not as popular, and Circlets are not as strong. Instead, main stat items (well, except for Mantle) take their place with 55% lane winrate each. The worst items to go for are Quelling Blade (48.98), Mantle of Intelligence (46.68%), Orb of Venom (46.92%), Gloves (47.89), Wraith Band (48.03%), Null Talisman (48.98%).

Offlane meta is not that different from Safelane, except the lane winrates are lower all across the board. Supports seem to be generally similar to Offlane, but with items like Smoke (56.22%), Mangoes (53.6%+), Blood Grenades (54.11%) and Magic Stick (54.95%) giving a significant boost.

Wait, but what about full starting builds?

Well, starting builds with low number of purchases are filtered out, and all remaining Safelane Support builds have a Grenade in them, with Offlane Support having a couple of Boots of Speed as a starting item, but otherwise the same picture. Still, the most successful build seems to be Grenade-Stick-Circlet-Branch-Tango-Wards for either role.

Best Sidelane Core builds are Wand + 3x Branches, with 2x Circlet + 3x Branches to follow. Even Midlane somehow still ends up following this trend, though 2x Circlets build is valued slightly more.

Consumables joined the chat

But wait, there's more!

Consumables stats section was born during my test period for a certain team during TI11. The basic idea initially was simple: looking at the specific players and their laning habits: how much regen do they get and how are they playing usually?

Now these stats are a part of the main reports, along with all the tier 1 tournaments. Though, this section is still experimental and it's only enabled after the event has concluded (processing consumables takes a while, compared to other sections). For the same reason Consumables are only enabled for ranked reports after the patch has ended.

This section is still in its early stage, and I'm still looking for ways to make it more useful outside of competitive setting, as I sense a great potential in it. I think it would also work well with some sort of "aggression" metric.

Observations

Since it takes a while to update Consumables stats, I will be using the last patch with a complete data in this section – 7.37e (at the moment 7.38 already concluded, but the data was not fully collected yet).

With a fellow friend Casual (analyst of Avulus team) we took a look at the lane winrate vs mean purchases of an item by a hero. To reduce the number of complicated operations, we used the main hero laning stats (not divided by role), but still filtered heroes with less than 5% of their matches on a certain role to hide irrelevant data.

The biggest problem with this data at the moment is the issue with stackable starting items (more about it in the next section), but it still gives a rough idea of certain trends.

More of these plots will be on my Patreon/Boosty/Private chat, but here are some of the more peculiar ones.

One peculiar trend: while for positions 1-3 there isn't a clear trendline for additional regen, there's a clear trend for both Tango and Salve for Supports. Mango has a somewhat similar picture, while Clarity doesn't give a more or less clear picture.

Cores do not seem to have any significant spread in this regard at all.

These charts are more of a proof of concept and I wouldn’t take it seriously for now. The most value of this feature still comes from the competitive part.

Problematic moments and points for improvement

One thing I was concerned about is the teams abusing Player laning, starting items and builds, as well as consumables stats during events. So these sections for Tier 1 events are not available without Early Access feature, until 2 weeks pass after the event is concluded.

Now what comes to the current implementation of these features, there are some problems with it. While I have plans to deal with it eventually, this is something worth talking about.

First of all, this section has much more "garbage" data, than one would want. Partially because of the replay parsing issues (e.g. Tome of Knowledge just appears out of nowhere), but as bigger problem are mass purchase spam, selling the item straight away. Ideally, I'd have to set up a sort of filter for that in the future.

Another problem is grouping very similar builds. I'm not entirely sure if it's worth doing, but it's at least worth considering.

It would be nice to have facets for these filters. And it is indeed very nice, however I don not want to include it for now.

The data source for starting items is the inventory snapshot from Stratz, limited only to the first six slots. This means that backpack slots are not available for use. This also means that multiple purchases of a stackable item (e.g. Tango) will blend together into a single purchase. I think I found a clever workaround, but we'll see how it goes.

When it comes to consumables, the pipeline to processing the match data takes a lot of resources and takes a very long time. As it's still an experimental feature, it won't matter as much. And in the meantime I will try to optimize it even further. And I will also consider merging consumables section with laning stats in some way.

But that's not all!

But that's not all! With this update there are some more cool things:

• A lot of optimizations to Stratz queries (for Immortal Meta Trends)

• Inline search filters now also support negation (add a minus before the word), as well as logical "and" (add "&" between two conditions, e.g. "core & offlane")

• Added Section Markers to show what's new/updated in the stats reports

• Ukrainian locale for the stats hub is 100% complete now

• Interactive map, metadata, courier icons (including bigger item icons) were all updated to support patch 7.38 (and 7.38b)

Closing thoughts and plans going forward

So that's about it. I spent way too much time on this feature, collecting the data and doing all sorts of optimizations behind the scene. Spending time on Dota things while not being able to give all my time to esports is a bit rough, but hopefully other features I'm working on will come sooner than in a couple of years.

Also I'm slowly migrating to Substack now. In addition to that, I'm trying to promote my Patreon/Boosty/Telegram Tribute a bit more and post more exclusive content and stuff to private discord/telegram chats for patrons.

I don't have much to share there (as I usually try my best to make everything public and available to everybody), but that's a nice way to support my work if you wish so, get direct support line for my APIs or get Early Preview access.

As for right now – that's about it. Not sure about what will be the next cool feature to make public, but there are a couple of promising prototypes: starting with MVP and ranking, ending with skill builds and Turbo Meta Trends.

We'll see how it goes!

The idea is simple: find a method to evaluate how “diverse” and “balanced” meta is in a report (aka tournament or patch). But how to do that? And what it might show you?

Russian version

I was experimenting with this topic for some time before TI, and decided to explore it more after completing the builds feature. Estimating balance level also went hand in hand with the idea of new teams hero diversity formula.

But to solve this riddle, you need to ask another question first.

What is ideal balance?

And what can you consider the ideal diversity?

To be completely honest, I didn’t think about this at all for a while, but it’s the most important thing to decide on.

Ideal meta balance is when every available hero was picked and banned, and the numbers of picks and bans for every hero are equal.

How to evaluate balance?

There were a couple of ideas regarding this even before the “ideal meta” was described, but my first real approach was based on calculating the maximum deviation from the median value. While it seemed to be a good starting point, it didn’t really describe the meta as much.

Unexpectedly, there was a better approach hidden in plain sight.

There is such thing called quartiles. These are the values that separate the data set in equal parts of 25%. In ideal universe with an ideal data set of random data between the 1st and 3rd quartiles you get two blocks that have 25% of data, or 50% of data in total. But here’s a catch: we don’t live in an ideal world!

So let’s take the number of hero picks. If we just take values of the quartiles, we might find, that there are some values in the first and the last blocks that are equal to these border values. So we end up having more than 50% of values between the 1st and the 3rd quartiles.

If we take the ideal balance, it will have 100% of values in this range, since they are all identical. And the closer our data set to the ideal balance, the more values we get in this range.

I am an amateur and don’t really know all the right terminology, so I’m not quite sure what to call it. But let’s call this value a “Core”.

How to calculate it?

There are some important details. First of all, the number of picks/bans can differ between reports, and there might also be some problems when calculating the Core as the number of total matches grows. For this I decided to use numbers of picks and bans relative to median, which you might have seen as “MP” and “MB” columns in the “Picks and bans” section. Rounding it up (and setting up our Q1 value to 1) we get a good description that should work for calculating the Core.

Second, there are three different values to go with: picks core, bans core and winrate core. And while picks and bans are relatively straight forward, winrate requires a more clever approach. For this instead of using [ Q1, Q3 ] range I went with [ m-0.5*M, m+0.5*M ], where m is median winrate in the dataset, and M is difference between the median and the maximum winrate (although it would be better to take the maximum delta without sign instead).

At first I wanted to get a single number to describe the balance level for every report, but decided to leave all three numbers and add an average balance rank instead. This makes it easier to see the whole picture: how balanced the meta was in terms of hero popularity, their winrates and bans. The average rank was calculated as a simple mean at first, but since then I updated the formula slightly. Now winrate balance matters most, then picks, then bans.

What about team diversity?

It’s a bit more complicated.

The approach is rather similar, but now we have two values.

The first is Core, aka how “balanced” picks of the team, or how many heroes were actually played (not counting outliers).

The second number is relation of the number of heroes picked to the theoretical maximum of heroes picked by the team. In other words, how many heroes the team could’ve picked by these number of matches played if every player was playing a unique hero every match.

Both numbers represent a different thing, so the end result is calculated as follows

(where Total is the total number of heroes available)

You can check out all the results in my stats hub.

The best examples are, in my opinion, last week at Immortal Rank report, Immortal Rank reports in general, as well as all reports regarding the last DPC season.

Interesting observations

It’s a bit hard to actually observe anything specific here, but it’s interesting to compare reports. In any case, there are some interesting nuances here.

First of all, it’s important to note, that the balance is estimated for the meta of a report. Meaning the “balance” may differ between tournaments. You should also note that picks and contest rate balance numbers are the least representative here since they are based on what people tend to favor more, and people tend to be biased and wrong.

It can be seen best when looking at the meta for the last week at Immortal Rank. Specifically when comparing Europe and China.

Europe has better winrate balance, but picks and contest rate numbers are skewed.

China’s situation in terms of winrates is worse (but it changes depending on the day). But picks and contest rate balance numbers are pretty high. It’s because in China people tend to play Random Draft in ranked matches, so the distribution is much more equal here.

And while Europe and China have more or less similar balance numbers, the average balance for all region gets as low as 68.6.

I also made two tables with the balance values: one for The International series and one for Immortal Rank reports (starting with 7.07).

Values in the tables below are slightly different from the final ones (because of the changes to calculation of average balance metric), but the relative position of reports pretty much didn’t change, so it works for getting a more or less decent understanding.

TI1 takes the first place, but mostly because it was the first: there weren’t many heroes available in the game, and there weren’t many matches played.

As for TI10, many people are biased towards it since it was the most recent TI, but it’s still one of the most balanced TIs and patches, along with TI9 (which is often confused with TI8). The most balanced tournament in the series, however, is probably TI7.

There’s also an interesting situation happening with the ranked reports.

The first thing you can notice is patch 7.19: it was the patch of The International 2018 (the 2nd worst TI), and it seems to be indeed one of the worst balanced patches.

Patch 7.29c seems to be the most balanced, and his brother 7.29d is the most balanced patch across the three longest recent patches (7.27d, 7.29d, 7.30e). Patch 7.30 is very close to them, so is 7.30d (the TI10 patch), but 7.30e dropped rather low. The reason for this is probably the meta of the patch being “figured out”.

Generally speaking, often the first versions of the patches (a/b letters or even without one) tend to be higher by “balance” then their later “brothers”.

And it’s also interesting to note that the most balanced patches are mostly ones that are recently released.

Closing words

Well, that’s the fun little experiment. It doesn’t have a good practical use, but it was fun!

Now you will be able to use beautiful numbers to complain about bad dota patches.

And that’s about it!

And here’s a mandatory reminder about my VK (ru), Telegram (ru), Twitter and discord. And the donations page, it really helps a lot. See ya!

For those who wasn’t there before, aside from constantly updating spectral.gg and working with teams PuckChamp and REDACTED, I’m also creating/updating a series of in-game guides called “Nerds Hero Builds”. And I’m also really into experiments with data in Dota (thanks to STRATZ for providing match data to make this possible).

This time I want to show you results of my long-running experiment, that started a bit more than a year ago (and eventually got delayed because of all the TI10 shenanigans): item builds based on item stats and progression trees.

Русская версия

You could’ve seen my in-game builds before: they all have “Nerds” in their name, and I’m constantly updating them (as well as adding new ones) since 2016, using stats and trends among the best hero players as a basis, as well as covering unorthodox ways to play heroes (like Morphling Support, Ion Cannon Luna, Magic Sniper or Crystal Maiden Core: another passion of mine).

I don’t always get enough time to improve on them, but this topic was always interesting to me. And my interest was just as big as my dissatisfaction with the existing tools

So here I will show you, what I’ve got, what was the problem and how I dealt with it.

The result

I’ll start with the thing that will be the most interesting for the biggest part of readers.

Let me present to you: the stats-based hero item builds section.

Yes, it has a lot of different numbers, but I tried to get everything as intuitive as possible, while leaving the information that will be the most valuable for decision making, without overloading the end user with it (or at least without drawing too much attention to the info).

Where are skill builds and starting items? Sadly, I just didn’t track starting items, consumables and skill builds before, so I don’t have this data now. I have plans to come back to this in future, but honestly I don’t see this as a high priority task.

Because in the end all of this (either starting items, consumables or skill builds) can be easily chosen on the fly. While the item build can help you understand the hero, it’s strong and weak points and understand how to skill it. (besides, there are more important things I have to work on)

Why are there so many numbers? I left everything that might be handy for decision making. I chose the stats that I was using myself, as well as things that were pointed out by high rank players I showed the earlier versions to.

Why there are two kinds of neutral items? The first neutrals block shows the best of the most common neutrals as a part of build overview. Bottom neutrals block shows all the most common choices.

Who is this for

• first of all it’s made for myself: this way it will be easier for me to update the in-game builds, which are…

• …made for nerds and very passionate players

• as well as pro players: while testing this feature I showed the final result to players of various teams, who pointed out multiple times that it increases efficiency of training

What’s so special about it

• you can see best (of the most commonly chosen) neutrals

• median timings of purchase, as well as winrate increases compared to games when the item wasn’t purchased

• for items with important timings: showing time frame as a reference, as well as winrate gradient

• it takes a note of alternative items that serve similar purpose and can replace each other (which also helps you compare items in a meaningful way, as well as understand why the hero builds a certain type of items and what for)

• it takes a note of situational items and their time of purchase as well

• and even shows possible swaps of order in which the items are purchased

• shows most commonly purchased early game items

• timescale that shows visually the order in which the items are purchased, as well as all potential choices a player might make

What it is NOT

• it’s not the full blown hero guide with skill builds, talents choices, starting items, etc: I don’t have data for these yet

• it’s not an interactive assistant that will guide you through the game

• and it’s not even an automatically updating build that changes depending on the heroes in your match: it wasn’t ever a purpose *, and besides it’s just a page on the internet lol

• it’s not a clone of Dota Plus or Dota2ProTracker (while the actual idea behind item builds trees and this feature took shape in one discussion with the D2PT dev we had a while back about how to approach item builds feature on D2PT)

* and a small footnote regarding “enemy dependent item builds”. While doing some experiments with the Immortal Rank data, I noticed that items stats don’t change significantly depending on a specific hero in the enemy team, except for lategame items.

Which made me realize: usually you buy core items for a hero to make it stronger by either amplifying the hero’s strengths or covering weaknesses. And once you’re done with them you get to lategame, where the real branching starts.

But my approach to this is simple enough: I want to make it more clear what choices make sense, but leave the final decision to the player — this is the idea behind alternative and situational items. It wouldn’t really be a good thing to just generate “a guide” to follow without putting any thought into it.

Where can I check this out

You can see this beauty in action in the Stats Hub on my website, in any report that has item build trees for roles (e.g. DPC 2021 reports, competitive meta and so on).

I would recommend checking out the Immortal Rank Last Week report: it always has just enough data and it’s always up-to-date: https://stats.spectral.gg/lrg2/?league=imm_ranked_meta_last_7&mod=items-builds

And if your first thought while looking at the example above was “it looks cool, but I didn’t understand a thing”: don’t worry. You can open any build and click on the “Explain this to me” button to see the detailed explanation of what’s happening.

And you can also read further: below I will explain how it all started and how it all works.

How it started

So for a long time I had a question: what can we learn from items stats? However, I wasn’t able to find a proper answer to it for a long time.

Items stats on Dotabuff only represent the endgame inventory, and this doesn’t really represent the real picture. Stratz, on the other hand, allows you to use time filters for specific heroes and represent purchase times, but the way the info is presented makes it hard to understand what’s where.

Eventually this led to creation of items stats section on my website (you can read about it here). This section was heavily inspired by a post in STRATZ blog about item timings. And this post is the reason why I implemented winrates depending on early/late time of purchase.

And while working with items I also decided to combine the data with another thing I love a lot: graphs, creating item build trees, representing item build progression visually.

Around that time the idea of “stats based item builds” was born. That’s why I added timings data, critical timings and manually grouped items into specific categories (partial items, early items, minor, major, etc).

The item build progression trees idea was in my head for a while, and around August of previous year it took shape, thanks to the discussion with the Dota 2 Pro Tracker developer.

Our approach to item builds ended up being rather similar (basically the same trees idea), but looking at the final result of “averaged” resulting builds based on the trees left me dissatisfied. Only with the actual result a major problem surfaced, and I started looking for ways to approach it.

The problem

The issue I was dissatisfied with sounds rather simple: the builds don’t have the full information, while also often containing additional “noise”, like items that sneak into the build without being an actual part of it, or small items, stuff like that.

The root of the problem lies deeper: the machine doesn’t “think” like a player when creating the build, and the end user doesn’t have much choice or understanding of what options are there. This problem is a part of Stratz builds, in-game Dota Plus builds and even D2PT.

It’s all while some items are separated into a special category (like boots, which might not be purchased or upgraded by some heroes), or rely heavily on metadata (like items cost and recipes), which might make it less universal.

What should we do?

Let’s take a step back.

Back when I started making in-game builds, I ended up with five major categories of items:

• early (everything that’s purchased before the main items)

• core (everything that’s important to use the hero to its full potential)

• situational early (items that you get along with/instead of early and core items)

• lategame (items that are not essential for the hero, purchased after core items)

• situational (everything else that might be purchased, depending on the situation)

These average builds usually end up showing only early, core and some of the lategame items (maybe all at the same time). While situational items often deserve more attention, since these additional choices reflect the actual nature of the hero and purchased items, while also making the player think.

These additional “branches” of decisions can be noticed with the item build based tree. This was the actual purpose behind adding this feature over a year ago.

But this leads to another problem: the tree can show multiple different routes to build a hero, it can be interactive, but even for an experienced player this might be hard to use and understand.

The solution is simple: we already have the item timings stats, as well as the “build tree”. We just need a way to go around this tree, taking these branches into account.

If only everything was this simple. But the general idea is pretty much this.

How it works

The first drafts of how it was supposed to look were just simple paper sketches and they are long gone by now. But these are the closest “electronic versions” of them.

And the first major problem that occurred is related to this plural “Builds” in the title. Initially I wanted to go through one tree for one hero, “splitting” it into multiple playstyles.

However, it wasn’t as easy. First, not every playstyle is equal in popularity, so some end up being filtered out in process. Second, playstyles are a spectrum, they might intersect and have completely different key timings, as well as item purchase order. Third, considering previous two things, finding and comparing two different builds, leaving only unique ones, just takes too long.

So I decided to go with a simpler approach: there is a separate build for every role (if the hero has at least 5% of matches played on the role). This covers the biggest branching factor of playing on different roles.

And for one role different playstyles may be covered by alternative and situational item choices.

On paper generating item build by traversing the tree sounds easy enough: you start in the earliest point and choose the item with the most purchases after the current. With this approach, however, you miss out on two major things: some items might swap their order of purchase (meaning one item might be purchased after the other or vice versa), and in some cases the item doesn’t have a specific purchase order, but appears in the build at some point.

This approach also leaves you without additional item choices, including “alternative” items (or “alts” for short: items that might be purchased instead a core item, not affecting the rest of the build) and “situational” (items that might be skipped altogether, without changing much).

So I ended up with something more complex: every of the nodes gets an “order” assigned, and potential “alts” and “situations” are marked accordingly, and if a match from the current “round” appears as a child in the next one, the number of matches is combined, while the pair of items is marked as possible “order swap”.

As for the commonly appearing items without a specific “order” — I added “favor points”: additional purchase number multipliers based on the item’s placements in previous rounds (the earlier a round happened in relation to the current one, the less valuable it will be).

The last issue is pointing out a specific point in time (or rather the build) when the “lategame” starts, since every hero has a different number of “core” items. And this becomes easy because of a simple observation: core items usually go along a very clear and unified path, and the branching starts once it’s time for “lategame”. So we just need to track down the item after which this branching starts: this will be the “going to late” point.

It all sounds quite boring, but the main point is simple: these tricks covered all the problems that caught my interest. The only unsolved part would be using additional items (not appeared in the build) and purchase times, but this can be easily covered by stats, that also include purchase timings.

Unsolved issues

Of course, there are still some unsolved issues left.

First: the builds needs data, the more = the better. Obviously, if there’s not enough data, you have to improvise. I ended up using “pure stats” of items by grouping them together by purchase timings. It somewhat helped compensate the problems for cases when there isn’t enough data.

Good example would be TI10 report and builds in it. In cases like this you will see a special warning about the build not having enough data.

Second: the builds don’t take metadata into account (like items recipes), so sometimes you might see items that are OBVIOUSLY combined, but their combined version doesn’t actually appear in the build. But the only thing to do for now is rely on the player’s understanding of the game.

Third: sometime situational items are shown as “alternatives” and vice versa. It might disappear naturally, but in most cases it’s enough to just think for a bit when spotting something suspicious (and generally treat alternative items as another version of situational).

The last one is a rare beast, but pops up from time to time. Some heroes are flexible and don’t have a specific playstyle, which leads to some wild results. But in a way it’s just a nature of these heroes.

Generally speaking (in my opinion) these issues are small enough and don’t affect the overall experience that much if you’re being thoughtful about it — but this was expected from the very beginning.

It doesn’t matter I’m not going to revisit these issues eventually, but they are not going to be a priority for me.

What’s cool about it

Aside from how cool it looks, the main thing that I love about it is the ability to understand how the hero is played and what to do with them in, all of that in a quick way.

When I was showing this to players of the (professional) teams I am working with, as well as various influencers, most of them immediately pointed this out (as well as the builds being pretty accurate).

The most important aspect of it is being able to understand how to play the hero by looking at alternative items. Using Spectre’s item build as an example: Falcon Blade is commonly used as an early stats item, as well as a farm item, since it gives a lot of mana regen to compensate for constant Spectral Dagger spam. Alternatives to Falcon are other items that are used to compliment the same sides of Spectre: Echo Sabre (tankiness, damage, mana regen), Soul Ring (tankiness, mana restore), Blade Mail (armor for tankiness, farm by using passive aspect of the item), Hand of Midas (maximize your farm, damage through attack speed).

And it also allows you to compare all alternatives at the same time. Combine it with reduction of visual noise, and you can concentrate on the most important things.

Displaying it this way also helps you understand the key timings of the hero, as well as power spikes.

Of course, there aren’t any explanations about specific item purchases against specific enemies, but I don’t think it’s needed: usually item builds rarely focused towards countering an opponent, but rather meant to compliment the hero and make them stronger. But displaying it all this way helps you reduce the number of options to choose from, speeding up and simplifying the decision making process.

As a nice bonus you get order swaps and situational items displayed with their approximate winrate shifts. This helps you see clearly what you get and what you lose by prioritizing one of the items or throwing another one into your build. And this, in turn, helps you understand in what cases this choice is worth it, and in which it’s not. And if you decide to go with it, how to minimize the losses.

As a side note, a list of things I’m aware of, but will probably figure out later:

• it’s not very mobile friendly

• neutral items appear twice: first time there are only the best ones, then as a full list (this was my intentional choice, but I’m not 100% sure about it)

• there aren’t starting items and skill builds (I don’t have the data for now, but I plan to add it once I’ll be able to)

So what now?

I have some small experiments left this year, but that’s probably it for this year. The only upcoming thing will be the yearly Spectral Report about what I’ve done throughout the year.

While the item builds don’t have the starting items and skill builds data, I already have some plans on fixing that. I want to create a brand new website subsection for my Nerds Hero Builds, where I will include all of my in-game builds, while also adding these data-based builds (and appending additional info to them). So I’ll make a slightly improved version of it rather soon.

Speaking of, in future this approach to builds might become a part of Dota 2 Pro Tracker (at least we were discussing this).

I also had some cool ideas on how to use the builds to create the power spikes charts — another idea that I was thinking about for quite some time. Current implementation of it already looks promising. However, it’s just a prototype, and I’m not sure when I’ll get back to it.

My main goal for now is to finally throw away the old junk of a code I was using and update everything, including the data schema, to add more cool stuff in future. Then we’ll see.

Closing words

That’s about it for now!

I’ll return soon enough, and will show you even more exciting stuff!

If you don’t want to miss it, you can follow me on Twitter or join my Discord (or you can also follow me on VK and Telegram if you understand Russian).

I often show sneak peeks of early stuff and experiments on my discord, so there will more than enough things to draw your attention.

And you can also donate here, this will be really helpful.

Until next time!

And did you ask yourself, what hero is the strongest in the right hands, and with what hero you can win even without any experience of playing it?

Or maybe what hero is objectively better than others?

How would one find the hidden OP hero?

Well, sit tight, and get some hot tea for yourself. We are going to the fascinating world of stats, where I’ll tell you about players’ experience.

Русская версия

You can also subscribe to my Twitter or join my Discord (or VK and Telegram if you’re Russian), where I talk about this kind of stuff a lot earlier, and also discuss dota stats, theorycraft some new features and monitor trends.

Introduction

You may have thought that all I did for the past couple of months was playing Genshin Impact and rewatching House M.D.

And, well, yeah, it’s pretty much most of what I do, but I also look into my discord chat a lot. And sometimes there are some interesting ideas that catch my eye, so I start implementing it so see what happens. In past two months I had some rather peculiar experiments, and today I would like to talk about one of them.

One idea I found interesting was a suggestion to estimate how different are winrates of heroes when played by different players: ones who have less matches on the hero and ones who have a lot of them.

It sounds interesting on paper: you can use this data to find heroes with higher skill ceiling, and find out their real power level. But here’s the main question: how would you calculate it?

Technical part about calculations

Now I will tell you about how I calculated this data, and if you’re not that interested in this kind of stuff just skip ahead until there are fancy pictures.

Among the things I could use there were, as usual, Immortal Rank data for the latest patch (7.28c). This data works great for this kind of experiments: there is a big enough dataset and the high mmer meta in general is pretty close to the competitive one.

But aside from the actual match data and players’ IDs I have no data at my disposal. It means that only matches happened in this timespan matter, regardless of how many games the player had on this hero in total.

On the other hand, it actually makes sense. If we want to make an objective estimate of the heroes’ power, historical data would affect the final result. So now we have the next question: how should we calculate the data?

It’s rather simple, too. First of all, we should find every player of every hero and calculate the number of matches they played.

The resulting set can be divided into groups of players. We have the most interest in players who had the most and the least games played.

This set would, however, also have a lot of players with a low match count, so we need to make some adjustments for our estimation: for example, all players with only one match are sent to the “low match number” group by default.

And the rest of the set can be divided into four equal parts. Players in the first quarter (before the first quantile, or Q1) will represent players who played less matches than most of the players. Those, who are in the last quarter (or have more than the third quantile matches, or Q3), on the other hand, will have more matches played.

With this approach we now know how to divide the players into groups. This, in turn, allows us to estimate their average win rates, and as a result we will have information about how often players, who had less or more matches on a hero, win. Exactly what we need.

So what can we learn from that?

First of all, we should point out the variables we have at our disposal, as well as what kind of data was used.

All the numbers in this text were based on the Immortal Rank Meta Report for patch 7.28c from the 7th of April. Of course, there will be a new massive patch after the 9th of April, and the meta will change significantly, but I don’t really want to wait for at least two weeks to get another big enouth dataset. And, well, most of the principles will still work for the new patch as well.

So what’s about the the variables?

Every hero has these groups of variables:

• Numbers of players with only 1 match played and more than one match

• Thresholds for every group of players (Q1 and Q3)

• Max number of matches played by a single player on the hero (for the whole timespan of the report)

• Groups of values for less experienced and more experienced players: average winrate, average number of matches on the hero by a single player in this timespan, total number of players in this group

• Winrate differences: gradient (how winrate changes with every additional game played — it’s not the most useful number at the moment, I’m still figuring out what to do with it) and difference (raw difference in winrates between more experienced and less experienced players)

You might ask, why exactly ranked matches? Why whole patch? Why not a week? Why not competitive?

The answer is simple: the size of the dataset. Last week meta reports have some fresh data, but they also have less matches, so these reports are more vulnerable to any random fluctuations from the least popular heroes. Data in the “biggest” patch report is the most stable.

And with competitive matches the answer is a bit more complicated. This feature is available for the competitive year report and even DPC Season 1 report. But, first of all, there is much less data (significantly less). But overall trends there are more or less similar to what you would see in the Immortal Rank report.

Most games played

What could we point out first? The obvious thing: on what hero players are usually played the most? By how much?

The first group in this sense is not that interesting: because of a big number of “one match players”, average number of games for less experienced players is usually somewhere around 1.25. But what’s interesting is the average number of matches for more experienced players.

There are two ways to estimate it: flat (take the average number of matches as is) and as a percentage of the hero’s total number of matches.

And just to be clear: charts below are all used with modified scales (so you could actually see the difference in some cases), so if the bar looks like zero, it’s not. Pay attention to the scale on the left.

And let’s start with something fun: flat numbers of max matches played by a single player on a single hero.

In top-10 by the max number of games on a hero by a player there are Naga Siren, Windranger, Venomancer, Visage, Pudge, Huskar, Phoenix, Broodmother, Pugna and Night Stalker. Their numbers are all somewhere between 200 and 250.

All of these heroes (except for Night Stalker and Pudge) are quite popular “cheese” picks, and also are quite popular among the boosters. Pudge doesn’t quite fit here, but doesn’t really surprise anyone: he was always loved by the players.

What’s interesting is Night Stalker: in patch 7.28c there was one Immortal player who played 444 matches. Why exactly Night Stalker? I have no idea.

The picture gets slightly different when looking at the numbers of matches relative to the total number of matches of the hero. In this version of the list there are still cheesy heroes, but there are more of less picked heroes, who are generally considered weak in current meta. Night Stalker doesn’t really stand out as much anymore, but I’m surprised to see Spectre and Brewmaster here. And on the highest spots you may see Lone Druid, Huskar, Visage and Chen players.

Among the heroes with the least flat values of max number of matches of a single players there are some out-of-meta heroes: Drow Ranger, Faceless Void, Sven, Omniknight, Viper, Meepo, Zeus, Necrophos, Silencer, Kunkka…

…but it gets more interesting if we’re going to look at the percentages relative to total games. Worst heroes here are, on contrary, mostly meta heroes. The worst one is Puck, who was played in 7.28c by almost everybody.

But if these were max matches, what’s with the average number of games for more experienced players?

There are also the same two ways to estimate, which produce two different results.

The best hero by the flat values is, undoubtedly, Visage, but after that the picture gets a bit more interesting. Unlike max number of matches, there aren’t that many cheese heroes, but a lot of complex ones instead: Broodmother, Techies, Tinker, Io, Earth Spirit, Dark Willow, Hoodwink, Morphling, Elder Titan.

Worst heroes by this criteria are not as interesting.

Flat numbers show a lot of out-of-meta core heroes.

And relative numbers show a bit more curious picture: there are, again, a lot of popular meta heroes.

But if in case of max number of matches it could be explained by constant bans and counter picks, there we see some new heroes (Shadow Shaman, Tusk, Lion), and the main reason for lowering the result might be the lower total pool of players.

Players’ success

Alright, so we now know, on what heroes players are playing the most, and on whom the least. But it’s not THAT interesting, isn’t it?

So here’s the cool chart to begin with: now every hero is a point on a scatter plot, where coordinates on X-axis are percentage of maximum matches relative to total number of matches of the heroes, and Y-axis are winrates of players who played the most matches on this hero.

And now let’s think, what can we do with the data?

First of all, we can find out average number of wins of both groups of players to then compare them with each other. With this we will find out what heroes are best and worst for more and less experienced players.

And we can get this plot! Horizontal position represents winrates of less experienced players, and vertical position represents more experienced players. Heroes closer to top right corner represents heroes who are better for both less and more experienced players.

And here we can spot some heroes who can help us point out a couple of interesting ideas.

First, winrates of more experienced players at high MMR can be used as an objective way to evaluate the hero’s power level. These players are playing the hero a lot, more than others, and not only do this quite often, but also know how to use the hero to its full potential.

Second, you can notice that strong meta heroes are usually closer to the top right corner. Some exceptions could be Tinker and Io (strong in current meta, but weak in less experienced hands), and the most top-right heroes are Enigma, Lycan, Razor, Nyx Assassin, Bane, Naga Siren, Beastmaster. A bit lower you can notice Night Stalker, Troll Warlord, Weaver, Shadow Shaman, Pugna, Luna and some others, who are rather popular in current meta.

And it means this kind of data represents current meta rather well, and also helps to point out heroes who are mistaken for being strong.

For example, closer to bottom left corner there are Pudge, Shadow Fiend, Grimstroke, Phantom Assassin, Keeper of the Light, Kunkka, Morphling, and a bit higher, but still close enough are Axe, Snapfire, Sven, Zeus, Outworld Devourer. Most of these heroes are still quite popular in both pubs and competitive games, but their winrates are weak at best, for both less and more experienced players.

Who should I spam?

Since we already found out how to evaluate power levels of the heroes, there’s only one question left: who should one spam?

What heroes are better with little to no experience, and what heroes are best with more experience? What heroes you should not play in ranked if you don’t know what you’re doing?

And stats can answer all those questions. And, as expected, meta affects it as well. So among the more “easy” heroes among the most effective ones would be strong meta heroes, and among the “worst” there are most likely to be the weakest heroes at the moment.

But first of all — who you definitely should not play without much experience?

The answer is quite obvious: complex heroes, like Lone Druid, Meepo, Tinker, Io, Keeper of the Light, Kunkka… Medusa? Bristleback? Pudge? What the hell are they doing here?

The reason is — you guessed it — the meta. These heroes had some problems finding their place in the game for a while now. In case of Tinker and Io the reason might be related to the need to know the exact reason why these heroes are busted and what makes them strong (specific playstyle and item build), and also to understand how to play them.

But it’s not all that simple with the best heroes as well. No hero has a winrate higher than 50% for less experienced players — meaning you should train these heroes at least a little bit before playing. Although, the closest heroes to 50% threshold are Abaddon, Razor, Nyx Assassin, Bounty Hunter, Shadow Shaman — which are rather straight forward. You might also spot some very strong meta heroes, like Lycan, Bane, Naga Siren, Night Stalker, Ancient Apparition.

And here’s an obvious question: if you need to learn to play anyway, who should you learn to play first? It’s probably some kind of Morphling, Grimstroke, maybe even Shadow Fiend? They are strong and scary…

…not.

Among the heroes who can’t get higher than 50% winrate even in good and experienced hands are Pudge, Shadow Fiend, Witch Doctor, Snapfire, Morphling, Axe, Phantom Assassin, Grimstroke, Lina. But what’s common for all these heroes?

The answer is simple: it’s a perception trap for a player (which is also commonly called “the fading bias effect”). Most of these heroes have some kind of skillshots, and if they don’t — they have some kind of “feel good” ability (like satisfying crit chance, or morph, or dunk).

And human mind is a fascinating thing: it usually remembers good moments (like getting a good crit, or landing a skillshot), but not bad ones (like losing a game). This may create a false impression that these heroes are good, even if in reality they are absolute trash.

But who should you actually train to enjoy some juicy MMR?

And once again there are some meta and cheese heroes. Interestingly, most of these heroes are usually called “situational”, but considering the high winrate with a high number of games, maybe Lycan, Razor, Enigma, Broodmother, Tinker, Naga Siren and others are not so situational after all?

“But where are Io and Clinkz”, you might ask. It’s quite simple, really: they are in the next 5 heroes after top-10. There are also traditionally good heroes in experienced hands: Lone Druid and Meepo. They, however, can’t be found among the best ones for a different reason: they are just too weak at the moment, so even the best players have a hard time getting a win with them.

Most complex heroes

And the last important value that you can look is the difference in winrates between the groups of players. In theory, this difference should mirror levels of complexity for heroes.

And here you can notice something weird. Among the most complex heroes there are Tinker, Io, Visage, Lone Druid, Lycan, Broodmother… But why Clinkz and Medusa?

How I pointed out earlier, with some heroes you need to understand what exactly makes them strong and how to play them to be successful. And, yes, the least popular heroes may pop up here for a similar reason.

Pretty close to the top there are Arc Warden, Vengeful Spirit, Huskar, Phantom Lancer, Techies.

And among the easiest heroes there are… Witch Doctor, Abaddon, Slardar, Lich, Skywrath Mage, Axe, Shadow Shaman, Juggernaut. The most unexpected heroes here are probably Shadow Demon and Snapfire.

Closing Words

My job was to tell you about what I did, how I did it and what stats may tell you. And, most importantly, this data is available from now on (more accurately, it was already available for two month now) in Immortal Ranked reports, as well as Competitive Meta reports, and weekly meta reports.

All the data in this text was based on 7.28c Immortal Ranked report, but most of the things apply for any other patch and competitive meta.

With this kind of stats it’s rather easy to spot when somebody “reinvents” a hero by finding a new way to play and win: for example, it happened with Tusk and Sven a month ago, when some players at high MMR started winning more than 65% of games on these heroes.

And in theory methods I used while calculating this data may be used to point out “more experienced” players to then explore their matches and item builds.

I listed the best and the worst heroes, and also made some useful plots to find a hero that suits you. And you can also periodically check the trends yourself and find the new imba before everybody else.

And you can also follow me on Twitter or join my Discord (or follow me on VK and Telegram if you’re Russian), where I write about various data anomalies, trends and other cool stuff when it happens. You can also donate here to support me and this kind of content, I would greatly appreciate that.

And remember: try not to get trapped by your perception and use stats, since stats don’t lie.

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 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.

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.

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%).

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%.

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%).

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.

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.

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!

Carry. Mid. Offlane. Support.

Long ago four roles lived together in harmony. Then, everything changed when the role stealer attacked.

Only the Flex Hero, master of all four roles, could stop them, but when the team needed him most, he vanished…

Dota is a weird game. Everything can work, everything can find its place. And even the most “boring” (for a common man) hero can have not one or even two, but many different playstyles.

If you’ve been following my posts (or seen them before), you may already know that I’m a big fan of obscure playstyles. And, well, it’s not just fun, but is also pretty handy: you may get very deep into playing just one hero, and find an answer for every role!

Today I’d like to discuss one rather interesting topic: why heroes can be played on different roles and which of them do it most often?

There won’t be much numbers and stats. Instead, I’ll share with you some of my knowledge I’ve got over the years of playing the game, researching meta trends and working with players, and note the most flexible heroes, based on positions stats for 6k+ MMR games in patch 7.27d.

Russian version

Core patterns

There are a couple of main reasons for a hero to be played on different roles:

a. The hero is too strong (in the current meta) and you don’t want to give it away to the enemy, so you’d get it for yourself instead and shift the hero to any free role.

It doesn’t happen as often in pubs, but is pretty common for competitive scene. The most important here is the reason why a hero can be considered “broken”: it may be a part of a very strong meta combo, or maybe it has an ability that is way too strong.

It may not be tied to the meta directly, too, if opponent loves to play a hero and is pretty successful with it. For example, if enemy loves to play Mars and has a high winrate, why not take it for youself?

It may work a bit differently in pubs, however. When a hero gets picked in the first draft stage too often and gets its role shifted (usually to a support role), then people start doing it in pubs as well.

b. The hero has a lot of playstyles, based on different skills and aspects of the hero

For example, Enchantress as a Core hero is usually built around Impetus, and Support is built around enchanted creeps and healing. Core Dark Willow is based around Shadow Realm damage, attack speed and elusiveness, coupled with the way to farm efficiently. Support Willow is based around utilizing her AoE control capabilities.

Arc Warden is a rather interesting example. This hero can be played as a classic hybrid core (Hand of Midas, Mjolnir and focused around Spark Wraith), as well as an aggressive physical damage core (based around Magnetic Field and Flux, built around Power Treads, Mask of Madness, Drum of Endurance and Desolator/Diffusal), an aggressive magic damage ganker (based around Flux range, Blink Dagger and Dagon), an utility Arc (focused on Magnetic Field and Meteor Hammer, usually played as an offlane core) or even a support (based around Magnetic Field, Drum of Endurance and Solar Crest).

Obviously, only one of these playstyles is the most accepted, but if you’ll look close enough and research matches of different Arc spammers at high MMR (and pros), you may notice, that all of these playstyles have a right to live — under specific conditions or in the right hands. Aside from that, some of these playstyles may be different to a point when they can be considered a different hero.

The most interesting and probably most common reason to switch, however, is adjacent roles.

Adjacent roles

Roles in Dota are not fixed. In fact, borders between them have been blurred to a point when roles can be considered a spectrum, rather than a set of five distinct positions.

Here’s when you can get to the first conclusion: heroes are often switching between two closest roles.

It’s most noticable with 1st and 2nd position, 3rd and 4th and between support roles.

The best examples of it may be Miracle- and MATUMBAMAN, who are famous for switching lanes and even heroes, or RAMZES666 and No[o]ne-. Another good example is Zai, who’s switching his roles pretty often: at different times and on different teams he could be an offlaner, while usually being a position 4, or a position 4, while usually playing as an offlaner.

Usually this kind of switch happens based on specific in-game situations and drafts, but more often the reason is a search for a better early game — in both competitive and pub games. So it’s not unusual to see “midlane” hero as a “carry” (and vice versa) if the hero’s early game may decide the direction for the whole match.

Supports are another story. If a couple of years ago there used to be a strict line between the 4th and 5th positions, it’s not so strict now, with free wards and cheaper support items. Heroes, that usually are played as position 4, mostly are potential midlaners (or cores in general), while position 5 heroes are usually played exclusively as supports.

And it’s still kind of hard to understand who is position 4 and who is 5 in a particular game, since the division is not as obvious as “offlane support is position 4” or “more wards means position 5”.

Another example of lane and role swaps is more likely to happen in pubs — lane swap between side lanes. It usually happens first when looking for a better laning matchup, but can become more common later. For example, it’s how Wraith King, Ursa and Chaos Knight started to be played as offlaners, and Bounty Hunter or Bristleback are pretty popular position one heroes at higher ranks and in competitive games. It’s most common for heroes who are strong during laning stage and in midgame, and who can pressure enemy safelane core.

This kind of metamorphosis can result in some other funny situations, when heroes (usually the ones who are strong in the current meta) switch through different roles to position 4 or even 5. At different times and in different patches this kind of thing happened to Wraith King (1 → 3 → 4), Sven (1 → 3 → 4), Chaos Knight (1 → 3 → 4), Monkey King (1/2 → 3 → 4), Invoker (2 → 3 → 4), Nyx Assassin (2 → 3 → 4/5). Sometimes it may to in the opposite way: it happened to Io (4 → 3 → 2/1) and Snapfire (4 → 5 → 4 → 3 → 2), but it’s much more rare and is pretty sensitive to any meta shifts.

Unique role in this regard is the middle lane core. Heroes here can be drastically different and can switch from any position. The main things here are the ability to win your lane (so you need to choose your hero for a specific matchup) and a hero’s ability to use the earned advantage (it can be a strong lategame potential, flash farming potential or gank potential).

And if the situation with safelane cores is already quite obvious and it’s mostly about matchup for them, it’s a bit more than that for others. Classic offlaners are mostly able to not lose their lane, and are also quite efficient at jungle farming. They can also make life much harder for their opponents, while feeling great throughout the game themselves. Good examples of this are Kunkka, Magnus, Axe (who is particularly interesting because he is able to skip waves between towers easily, which makes the lane much harder for enemy).

And since most of the support heroes are pretty strong laners by themselves, for them the main reason to switch their role is some kind of change that allowes them to scale better into the lategame.

Good examples of this are Dark Willow (who was pretty popular mid hero after being released, and became popular again after she got an Aghanim’s upgrade), Earth Spirit (who was pretty popular choice for midlane ever since his release because of his gank potential), Snapfire (after the hero’s release key factors were great base damage, efficient flash farming and lategame potential with the Lil Shredder level 20 talent, and came an Aghanim’s scepter), Dazzle (after his Poison Touch rework, addition of Bad Juju and Guardian Greaves change), Winter Wyvern (at first — because of her aghs, then because of added possibility to attack heroes under the Winter’s Curse effect).

In general, “carry potential” is not as important for a mid hero: most of the relevant mid heroes are usually played as something between core and support.

Many of transitions like this are not as typical for pubs. People here are more likely to stick to competitive meta and pick “for sure”, so their choices are more conservative. Often this kind of trends — be it playing on a new position or a new build for Anti-Mage — are easy to notice when watching high rank matches (7–8k+ MMR), but not as obvious when you’re looking at the “bigger picture” or even 6k+ MMR trends.

My most favorite example of this is 2B and Support Invoker. 2B is a 7k+ MMR player and streamer, who’s famous for constantly playing traditionally core heroes as supports and is quite successful at it. Often he’s being unfairly criticized for it by his team, and when he started playin Support Invoker, their reaction wasn’t any different from it when he played support Troll or Anti-Mage. And yet other players started noticing this kind of playstyle.

Support Invoker’s popularity wasn’t higher than 10% compared to the hero’s other roles at 6k+ MMR, but everything changed when YapzOr played it in an official match. After that part of Invoker’s games as a support skyrocketed to 76% and even now it still remains stable at ~12%. And it was pretty similar story for Ember Spirit and Sniper as supports.

Flex heroes

But what heroes are the most commonly accepted “flex picks” at the moment? Who can be picked for different roles and switch between them?

To find out I used positions overview for patch 7.27d from my ranked reports based on matches of 6k+ MMR players.

My main criteria for selecting heroes was having at least 10% of games on a non-primary role. It doesn’t mean that hero-role pairs with less than 10% are not significant — even 2–3% of games are already significant enough to pay close attention to a hero, but we are looking for the most common ones. In some specific cases I allowed myself to include heroes who are below this line, based on my experience from previous patches or observations from competitive matches/scrims.

Heroes are divided in 5 tiers:

• Tier 5: two roles, one playstyle

• Tier 4: two roles and different playstyles

• Tier 3: two roles with closer distribution of matches or three roles (that may not be represented in stats as much)

• Tier 2: more noticable distribution between three roles or four roles

• Tier 1: five roles

It’s also worth to point out: I didn’t pay attention to winrates as much. Aside from a few exceptions, win rates of different roles for one hero are pretty much the same, and win rates in general in this case is more about understanding the hero and its playstyle on a role better. Besides, if our goal is to find the most accepted flex heroes, win rates have the least importance.

The last thing I’d like to point out: support role stats is not separated because of technical reasons. So it may be both position 4 or position 5. I marked support positions mostly based on my observations, but support roles are so close to each other, that you can consider them one role.

Tier 5

Heroes, who stand out the most in terms of two position divisions, without changes in playstyles, are:

Here we can already notice the same transitions I talked about before. The most common pairs of roles are safelane and midlane cores, as well as offlane and support. The only two interesting cases are Lone Druid and Dragon Knight.

Lone Druid is traditionally coming to offlane, and yet his playstyle on all three core roles is pretty much the same, so you can say that LD is surprisingly stable.

DK is, however, the closest to cross the line of two different playstyles.

Tier 4

And here, where playstyles for two different roles diverge enough, you can see a pattern: the division often happens between a core and a support roles (the only exception is Death Prophet, but historically she became a common offlane core through being played as a position 4 first). The difference can be easily explained: usually these heroes are required to do different things on these positions, and are more likely to change both their item build and playstyle based on a different amount of resources they have.

The best examples of it are Lina, Mirana and Abaddon. In case of the first two, one build is based around cheap and efficient items (Tranquil Boots, Mekansm, Eul’s Scepter), spell damage (aghs, lens), auras and surviveability, while the other is based around attack damage (Monkey King Bar, Daedalus, Manta Style).

Abaddon’s builds are pretty similar, but different enough: support build is cheaper in general, and the game is mostly built around surviving, healing allies and buffing the team; offlane build has a bit more focus on attack damage, team auras and is based around the idea of being a tanky frontliner.

This tier is also the place where you start noticing significantly more different roles in a pair: midlane core and supports. It’s hardly a coincidence: many common mid picks have been supports back in the day (or became supports today), and usually we are talking specifically about position 4. The reason for it may be better lategame potential and greediness on one hand, and strong spells, coupled with meta, on the other. It also makes it clear: most midlane cores can become supports (and vice versa) if you’re creative enough.

Tier 3

In this tier we start to see many heroes, who are rarely played on some positions in pubs, but can be noticed on the highest level of ranked matches, in competitive games and in teams’ training matches.

One particularly interesting example of it is Mars. This hero is not as common for midlane anymore, and the support Mars trend didn’t get enough momentum yet to become noticable in stats. However, in the last 6 month more and more players started to pay attention to this hero as a support, and now at least four different teams are training strategies that have Mars as a position 5 support. Maybe we are going to see it in a competitive game soon, who knows?

Arc Warden is in a similar situation. Self has a very clear distinction between safelane and midlane, and it’s also one of the heroes who has different winrate for different roles (48% for sidelanes and 54% for midlane), but some teams and pub players are seeing it as an offlane or even support hero. There are a couple of reasons for it: first, he has a strong gank potential, second, he has a strong combo with Drum of Endurance and Solar Crest, and third, the Magnetic Field. But even though players were looking at the hero on these roles for a while now (especially in SEA and China), it still remains a niche pick that is hard to get going.

Enchantress was considered equally good as an offlane core and as a support for a long time, but not since the hero got changed last time. Today she’s almost exclusively played as a support — even though position 4 and position 5 playstyles may be drastically different for her — and popularity of other roles decreased significantly.

For other heroes roles division isn’t that big of a secret, but with higher skill level (or in team matches) it’s more common to see these heroes on more “strange” roles. Usually there are specific reasons behind it.

For example, for Dazzle it’s his strength on a one versus one lanes, as well as flash farming (given a good start) and efficiency of team items in his hands. In case of midlane Luna people usually consider magic damage build (aka “Orbital Cannon Luna”). Bounty Hunter and Timbersaw are pretty common niche carries in general, but they require understanding the game on a decent level. Doom was always popular as an answer to enemy Chen picks.

Honorable Mentions

And before we go further, I’d like to note some heroes, who were played on different roles before, but these playstyles are either forgotten, or lost their popularity.

• Sven Support was popular for a while after a change to his third ability (when it was physical damage shield), but after it got changed back, it was forgotten.

• Morphling Support (aka “Stun Gun Morphling”) is probably as old as Dota itself. I can’t recall it being very popular at any time, but it was always working and people still tend to remember it. And, well, it’s not surprising: who wouldn’t love to throw some 4 seconds stuns to enemies?

• Slardar Safelane/Midlane is not really forgotten, and it is picked occasionally at higher ranks, but not in general. The main playstyle is pretty much the same as for offlane in this case.

• Lifestealer Offlane/Support was popular for about a week, when Infest started to give additional HP to allies, and aghs was changed to upgrade Rage.

• Axe Support is played in about 5% of Axe’s games and was popping in high MMR game for a long time now. Who knows, maybe we’ll see it in a competitive game one day?

• Skywrath Midlane used to be popular, but now he’s picked only in 5% of Sky’s games, when stars align.

• Ancient Apparition Midlane is in the same situation as Skywrath. AA was a popular choice for midlane before, but now it’s considered to be more of an exotic choice. He’s still played occasionally, and AA’s new aghs allow him to scale better into the lategame, but usually it requires way too specific conditions.

• Io Safelane/Midlane Core is not as wild anymore, and Ana, who made it popular, is nowhere to be found. But this playstyle is still used by 8k+ MMR boosters.

• Shadow Demon Offlane was never really popular, but was leaking to competitive matches, usually played by iceiceice and 9pasha.

• Naga Offlane/Support is a weird one. Naga always goes through some sort of a cycle: in one patch she’s usually played as a core, in the other — mostly as a support. Lately utility Naga is not as good, but it’s unique enough and deserves a mention.

• Dark Willow is almost never played on any role, but support. But it wasn’t always like that: she used to be a Midlane and Offlane hero as well.

I can list even more exotic stuff: support Troll, mid Underlord, mid Earthshaker, support Clinkz, even support Brood. But I guess if we’re going to remember all the exotic stuff, this list will become infinite.

Tier 2

This category is the first one to include my favorite type of heroes: “Panic picks”. Gone to kitchen, came back and there’s a game going on and you need to pick something quickly? You’re playing in party and still haven’t decided who’s playing what role? You can always pick one of these heroes! They have a lot of different playstyles, can be played on various roles, and you can even get more out of them if you know what you’re doing.

The most questionable one is probably Venomancer. Even though he has a distinct division between offlane core and support, it’s a clear division between pubs and pro games — what’s up with the other core roles? The answer is simple: almost every single patch, many teams are actively seeking for a way to add this hero into their drafts, switching him from role to role, training Veno strats in their scrims. The reason for this is him being a disgusting hero to play against. The hero itself may not be as strong, but playing against him is very uncomfortable, and the vision and sieging that Veno provides alongside this mental warfare is always valued in a draft.

The similar reason is behind Nature’s Prophet and Puck. Their playstyles are almost identical on every role, and they are frequently picked, just like veno, because they are uncomfortable to play against. These two are very efficient at split push, elusive and have high base damage values, which can make the laning stage unbearable.

You can also note Viper, who’s on the list for a similar reason. He’s often played as a safelane core as well (but it’s usually more about a lane swap, rather than about role switch). Silencer, like Viper, is able to become a strong additional core for his team, if played as a support. The only detail to note for core Silencer is priority of offlane over safelane: it’s most likely related to lane swaps and early aggression against enemy safelaner.

Snapfire stands out a bit. She is one of the heroes who has different win rates on different roles: 48% as a support and 41–42% on every other role. It’s hard to say what is the exact reason behind it. I think it has more to do with people’s understanding of the hero on these roles: Abed and Topson started the core Snapfire trend and were pretty successful with it in pubs. However, it could’ve been purely because of surprise factor, and once people got used to it, win rates fell off.

The most interesting heroes on the list are probably Chaos Knight and Lycan. Both of these are pretty strong laners, and both of them are rarely seen as supports in ranked pubs. It’s not rare, however, for higher level of play: more and more people are starting to consider these heroes as position 3–4 ones. The reasons for it are their aghanim upgrades, overall utility and ability to either sustain under pressure early on or play aggressively. And while CK as a support popped up years and patches ago, Lycan is quite new to this thing: only after he got the previous aghs upgrade he started to gain popularity as an utility core.

Tier 1

And who are the most flexible heroes? Who can be played on every role?

It’s not hard to notice, that most of them (like some of tier 2 heroes) are quite popular among boosters.

For a common man all these heroes are mainly supports or midlaners. But there’s more to it.

The thing they all have in common is their ability to do pretty much everything. Every each of these heroes has some kind of utility (a reliable stun, an aura, a vision and scout spell, or even everything combined), push potential (and you don’t even need to attack by yourself), surviveability (be it good stats, a spell for tankiness, an escape spell or even movement speed).

It clearly shows, that carry doesn’t need to be “classic”: Lechrac, Pugna or Visage are often played as safelane cores (if they are meta) in competitive games, and do their job — which is surviving and destroying enemy buildings — quite well.

(These aren’t the only things a carry should do in a game: you can be a carry in a lot of different ways; it is, however, the most important thing for these specific heroes to do)

The only hero that stands out from this list is Vengeful Spirit, a hero whose popularity across roles isn’t reflected in these statistics; However, the legends of Carry Venge are as old as the hero itself, and there are few people that still hold doubts about her viability as a midlane or safelane pick at this point.

It’s a bit more interesting with offlane: this playstyle gained popularity after the controllable illusion on death was implemented. After that, Venge, who already had pretty useful spells and great stats, suddenly turned into a strong tank-initiator, who can build all sorts of auras and die in teamfights without fear, while still being useful for the team with all the auras and natural tankiness of the illusion.

Closing words

Of course, flexibility of heroes is a relative thing, and sometimes you may notice completely different playstyles through item builds for the same role (we’ll talk about it later, I promise).

But please, don’t think that just because a hero was mentioned it can’t be played on a different role, or that any deviation from the “standard” is going to lose you the game — you can see in this piece how successful innovation still is in this game.

As I already mentioned, roles are a spectrum, and Dota is a weird game, where everything can work.

And that’s about it! While I’m working on new content and stats hub updates, you may want to subscribe to me on Twitter, YouTube or join my Discord (or even subscribe at VK and Telegram if you’re Russian).

You may read the whole thing, about the idea and implementation, or just straight up skip to the last part with some observations.

Russian version of this text can be found here.

The Idea

Initial version of the idea was suggested by Fazy on my Discord. But his initial suggestion was slightly different: winrates of heroes playing “from behind”.

The idea raised some questions straight ahead. What does it actually mean for a hero to be “behind”? How can we define the advantage? How can we detect it? However, I definitely liked the idea, and I was curious, if I’ll be able to actually implement it.

Boring technical part

The concept level of the idea looked fairly easy: just take the networth difference data and detect, who was behind in networth (aka enemy had higher networth) for a longer period of time, while also having a some kind of “tie” threshold (e.g. 10%). And for you to understand why exactly it was close to impossible, I need to explain, how exactly the “reports” work and why I don’t really want to add new features right now.

Back when I started making “reports”, I just wanteed to make a bunch of scripts, just a simple toolset, to automatically gather match data, saving only the data I need to local databases, and then “cooked” the complete “report” — an all-in-one file that would have all the aggregated and calculated data. Initially I thought it will be just a bunch of simple one-time-use scripts to make some Reddit posts about the tournaments and static HTML pages. Something went wrong.

In the end I didn’t have neighther Markdown module for reddit nor static HTML output. But I was still impleementing new features for the Web View, as well as all the other scripts.

Bear with me, it’s all boring, but it’s important to understand the limitations ahead.

So, all the things I explained earlier, led to these problems:

• The code was quickly thrown together and it wasn’t initially planned to be anything bigger, so it’s hard to support

• Changing the structure of a database is fairly easy. The problems started with multiple independent DBs, mostly because of the way the scripts were written. In reality the DB structure for a report is somewhat fixed

• Every report file (due to the way it was processed initially to geneerate simple static pages) is loaded wholly into the memory while processing web view, and well… Sometimes, when the report structure is big and complex, the memory may just end. It made complex data somewhat a taboo for some reports — and it may be easier to just rewrite everything from scratch (and I’m working on exactly that for a while now)

The latter problem got partly solved (and I may write about all the tweaks related to it), but it’s still a significant problem that can’t be ignored.

Slightly less boring practical part

With all the problems I listed just a bit earlier, we practically have these things to consider:

• We can’t change thee code structure or logic too much (otherwise this shitcastle may just fall)

• Database structure is practically fixed, so we need to work with something we already have

• The amount of data geneerated for the report should be limited in size

But what data do we have at out disposal to make any assumtions about the game state and calculate outcomes? My first thought was about “stomp” and “comeback” values, but those don’t even have any faction related information, so they are practically unusable. But then I noticed another field that was leeft unused for a long time: “lane efficiency”.

It’s practically the same as “efficiency” value in Opendota’s match analysis, and it basically says what percentage of the maximum available lane gold (passive and from lane creeps) the hero was able to accumulate to minute 10. This metric always looked interesting to me on paper, but I couldn’t find a good way to use it.

And this is how the idea developed: what if we will decide, if a team was behind or not, by looking at lane outcome? Lane lost = behind, lane won = ahead. And I already had an urge to play with lane stats, so why the hell not? Now the only thing we need is to make a test query, that would combine the data together.

In practice, we have lane stats, calculated from gold at minute 10 mark.

But comparing teams isn’t that useful, so why won’t we compare lane opponents instead?

But what do we do with Supports then? My initial thought was using average efficiency value of Core heroes on the same team as the support hero and compare it to an enemy support — practically using the team’s overall efficiency to evaluate supports. But this approach led to massive increase in query execution time when having a massive number of matches (e.g. meta or ranked reports), so this solution wouldn’t fit. This led to another thought: we may be able to use maximum efficiency value on the lane, so supports will be evaluated using their lane companion. At first I had some doubts about this solution, but it actually turned out to be pretty good in practice.

The only thing remaining is building the test tool to execute the query and look at the data, and…

Nice.

Now I just need to integrate it into reports and make a web view for it… and that’s where things became a bit more complicated: important data that gave laning stats context, was a bit too much: test report for patch 7.27d had 8x increase in size! It doesn’t look good. After “wrapping” the data in a nicee way, restructuring data and compressing it, the size change dropped down to 2.5x, which is still a noticable change.

Observations

So, that’s about it with the boring tech part. What do we have now?

I used available to me 5.8k+ MMR immortal rank data and generated a report based on it to look a bit more into laning data. Technically I could’ve used competitive data, but the overall trends are pretty much the same, while having much less available data.

First of all, I’d like to explain the table:

• Matches: basically the hero’s matches, but not necessarily. There may be slight differences from matches values in other tablees. These differences may occur either because of lack of detailed analysis, or because of lack of lane opponent in a certain game (this happens sometimes)

• Lane win rate: literally lane win rate. It works pretty much the same way you would expect from a regular winrate, but the main thing to keep in mind is that we may get a tie here. Tied match is considered a “half” of a win

• Rank: just likee with other sections that have rank, it’s some sort of “small winrate” sort. In this case parameters used for rank calculation are hero’s lane win rate, overall number of matches, average advantage (and disadvantage) on won (and lost) lane, as well as winrates for various lane outcomes.

• Average lane advantage: basically average difference in efficiency. This block has two values — average advantage on a won lane and average disadvantage on a lost lane.

• Categories for lane outcomes (won, tie, lost), every of which has two values: Ratio, which shows a ratio of this lane outcome matches vs total number of matches, and Games won, which shows winrate with this lane outcome

Now it’s time to talk about the cool stuff I noticed.

First thing that I saw is how surprisingly well this support solution fit into the puzzle. I evaluate most of the data based on my own experience in games, so I may be biased, but everything I saw was pretty much answering to all of my expectations.

What’s also interesting is the effect of using only one value, pretty much tied to gold earned, is practically invisible. I guess, it’s mostly because any of factors that may be able to indicate the lane outcome also affect the difference gold earned: kills, deaths, tower kills, ganks… There were only three cases that raised some concerns.

The first one was Alchemist. I was worried that he will be the winner of every single matchup — and yeah, he still has pretty high lane win rate, and yet… His values aren’t actually as high, as I expected. Partially it’s because he starts to accelerate significantly only with bounty runes spawns, stacks farm and, well, networth difference in general becomes more noticable only after the laning stage. And what’s even more important — there are some additional metrics, like winrate while won or lost lane, as well as average advantage and disadvantage, that give his winrate a better context. These also show that he’s not even the best early game gold earner, and his winrate after lost lane is 14th from the last place.

The second case was winning the lane by switching to jungle. But come to think of it, not many heroes are actually capable of outfarming lane opponent by going to jungle, while also not completely abandoning the lane (otherwise they would be detected as junglers). Not many heroes are even capable of efficiently farming jungle during early minutes, especially since neutrals give less gold early on. So, I think it’s fine to go with the assumption, that if a hero is capable of outfarming the opponent by switching to nearby jungle, then this hero pretty much won the lane.

The last questionable case, that actually caught me thinking for a while is “lost lane that was actually won” — the 1 vs 3 lane scenarios, when a hero gets focused. I had this question asked a couple of times, and even got confused for a moment here: what should we do with these lanes? And the answer came with the question: how is it even possible to “win” the lane while losing it? The lane is still lost, no mater how you call it. And to eevaluate the impact of the lost lanee we can use additional parameters, like win percentage while the lane was lost.

Best laners and snowball heroes

We can look at the heroes with the highest lane winrates straight ahead, and most of them are midlane heroes, with rare exceptions: Broodmother (74.32%), Alchemist (72.82%), Templar Assassin (68.55%), Viper (68.32%), Timbersaw (64.65% — can be played as a sidelane core), Doom (63.90% — not a mid hero, but can be occasionally picked as mid), Sven (63.26% — primarily safelane), Shadow Fiend (63.10%), Lina (62.81% — 40% mid, but mostly support), Bristleback (59.74% — occasionally can be core on any lane, primarily sidelanes), Nature’s Prophet (59.59% — also played in offlane), Meepo (58.62%).

It’s not very surprising, considering most of mid heroes are going to mid specifically because they are capable of either winning a lane or accelerating in farm by getting solo experience and a good start. But it also shows overall “brawling” nature of sidelanes: most sidelane heroes are usually forced to fight or react, and it’s rather common for safelane cores to have slightly less last hits at minute 10 mark as well.

While most supports and safelane cores have high winrates from won lane, but low or at least average winrates on lanes, most midlane cores have the opposite situation. It also clearly shows a thing many players had a feeling about for a long time: it’s easier to win a game from midlane,

The main trend that you can notice is most heroes at the top are also capable of efficiently farming jungle early on, which is the source of their rising lane advantage. It’s not that surprising by itself, but it’s surprising to see some unexpected picture shifts.

For instance, Broodmotheer has a significant lead over Alchemist in both lane win rate percentagee and average advantage. Even average disadvantage on lost lane for Alchemist is significantly higher than for most other heroes, and his winrates with multiple lane scenarios are lower across the board. It may point at both the hero being rather weak and the fact that he has a hard timee coming back from lost lane scenarios.

Another hero that is somewhat similar to Alchemist is Doom: he’s also capable of creating money out of thin air. However, it’s not his only secret: Doom’s stats look significantly better than Alche’s.

A hero that became an interesting observation is Lina, who is here at the top. And while she’s capable of being a midlane core, it’s more common to see her as a support hero these days (she is played as a midlane core in only 40.9% of cases), so she is practically a support hero. However, her winrates, especially the one on lost lane, are lower than aveerage.

It’s also makes Lina similar to Bristleback. This beast is not the most common guest on the middle lane, but has a fairly high lanee win rate, while having the worst winrate from lost lane: 22.1%.

A bit lower at the top wee can see three heroes that almost never go to midlane these days: Bloodseeker (primarily an offlane hero these days), Juggernaut (primarily a carry) and Enchantress (she is mostly picked as a position 5 support, but can be seen as a position 4 support as well, or even midlane/offlane core). These heroes are not the strongest on their roles (Bristleback/Timbersaw, Sven and Lina respectively are a bit ahead), but stability of their stats is more noticable: generally speaking their winrates with every lane outcome are just, well, average. What’s even more curious, they are not early junglers (except for Enchantress, but it’s usually the case for support-ish Ench, and in this case her own farm, acquired in jungle, doesn’t matter).

I’m generally surprised with this list of heroes: it fits almost perfectly into my picture of the game. Some of the heroes that are at the top with lane win rate are practically created for lane domination. Others are pretty popular in current meta exactly because they are capable of dominating the lane, defininfg the trend.

Winning with the lost lane

Going even deeper into lane outcome winrates discussion: there’s not a single hero who has positive (50%+) winrate with the lane lost. In other words, losing your lane is going to increase your chances to lose, regardless of the hero, and it’s better to try to at least draw the lane if possible (e.g. by disrupting enemy’s farm as well if you can’t farm yourself, or acting like you’re actually winning your lane, forcing enemies to back out).

But, obviously, some heroes don’t need to win the lane as much as others. And, there’s indeed some interesting spread of values.

Let’s define context first: Invoker has the highest winrate with the lane lost (39.69%), Bristleback has the lowest (22.1%), and the median is somewhere around 33.8%.

I expected to see heroes like Spectre, Anti-Mage or Terrorblade — safelane cores that usually lose their lane and catch up in the jungle — at the top. But, surprisingly, most of top-10 heroes here are actually supports and offlaners. And the top five are Invoker (39.69%, primarily midlane, can be a position 4 support), Chen (39.66%, almost exclusively a support), Visage (39.02%, can be played as safelane, midlane and offlane core, as well as position 4 support), Clinkz (38.87%, almost exclusively safelane/midlane core), Magnus (38.71%, mostly a midlane core, but can be seen as an offlane core and even as a support).

More interesting, however, is the list of heroes with the lowest winrates on lost lane: aside from Bristleback down here at the bottom you can see Gyrocopter (25%), Wraith King (26.13%), Pudge (26.99%), Lina (27.46%), Omniknight (27.46%), Kunkka (27.5%), Enchantress (27.51%). It’s also worth to note that most of the heroes that have relatively high winrates with lane won can be found here, including some traditional autoattacking type carries (like Monkey King, Lifestealer). It can signify their need to win the lane and snowball from there.

Many heroes, however, don’t win a lot of lanes, but have a high winrate when the lane is tied. It’s not rare to see a hero with a positive winrate on a tied lane, and here at the top we can find a lot of currently popular meta heroes, that may not be able to win the lane as much, but at least are capable of forcing a draw. At the top of the list you can notice Clinkz, Faceless Void, Spectre, Chen, Enigma, Lone Druid.

Strongest laners for every role

But if the lane outcome is so important, what should I play then?

Worry not! Just for that I’ll list some best laners for different positions.

I’ll sort the list by rank and limit it to 6 best heroes for every role, that also have lane win rate higher than 51%.

• Carry: Sven (63.26%), Juggernaut (58.19%), Huskar (57.11%), Lycan (55.71%), Lifestealer (54.96%), Terrorblade (53.71%)

• Midlane: Broodmother (74.32%), Alchemist (72.82%), Templar Assassin (68.55%), Viper (68.32%), Timbersaw (64.65%), Shadow Fiend (63.10%)

• Offlane: Timbersaw (64.65%), Doom (63.9%), Bristleback (59.74%), Nature’s Prophet (59.59%), Bloodseeker (58.23%), Axe (57.52%)

• Support 4: Lina (62.81%), Enchantress (57.98%), Leshrac (55.38%), Tusk (52.58%), Venomancer (52.55%)

• Hard Support: Enchantress (57.98%), Venomancer (52.55%), Undying (52.26%), Lich (52.06%), Treant Protector (51.59%), Io (51.53%)

And if we’re gonna talk about the strongest ones, why not look about the worst ones as well!

Here are the list of 5 worst heroes for every position to lane with (unique for every position)

• Carry: Spectre (37.58%), Faceless Void (42.61%), Wraith King (47.95%, rank 10), Lone Druid (47.27%), Drow Ranger (45.47%)

• Midlane: Visage (36.75%), Sniper (46.36%, rank 0.8), Tinker (37.43%), Puck (39.06%), Ember Spirit (37.78%)

• Offlane: Omniknight (45.33%), Pudge (44.09%), Abaddon (45.46%), Brewmaster (42.96%), Pangolier (41.24%)

• Support 4: Techies (41.10%), Enigma (45.67%), Nyx Assassin (45.24%), Earthshaker (44.51%), Tiny (45.18%)

• Hard Support: Keeper of the Light (43.23%), Dazzle (48%, rank 5), Grimstroke (46.6%), Shadow Demon (45.88%), Winter Wyvern (44.34%)

And everything here looks obvious and nice: we already saw most of it in our games after all. But the most important thing here is this observation: while winning a lane is important and may make your game significantly easier, it’s not the most important thing. For instance, among the “weakest” carry heroes there are three (out of five), that are relatively popular in current meta and are pretty strong. So the key thing here is the bad lane is not the end of the world, even though it’s worth avoiding if possible.

Closing words

Generally speaking, the main thing you shouldn’t do is picking a hero that will lose the lane. Ideally, if you already have a bad lane matchup, it’s better to force a draw or at least act like you won the lane, so nobody will suspect a thing.

But how do I pick a hero for a lane matchup? For this specific case I enabled the laning component for every immortal ranked trends report starting with 7.27, as competitive season report and meta reports. Not only you can research overall laning stats for all the heroes, you can also look for a better or worse matchup against specific hero as well.

Share your findings and in the Spectral Discord, on Twitter (@ileamare) or send it to my VK page.

And you can also find the 7.27d laning stats here: https://stats.spectral.gg/lrg2/?league=imm_ranked_727d&mod=heroes-laning

And, well, try not to lose your lanes, and even win most of them!