67 replies to this topic

[Jylpah]

Hello Fellow Tankers,

 

I have set up a new Blitz web site BlitzAnalysiz.com - a site focused on in-depth analysis of game statistics. From the site you can find answers to questions like:

1. “What is the most over-powered tank in the game?"
2. “What is the most played tank in the game?"
3. “How is players’ WR distributed over the tiers?".

 

You won't find individual player statistics there, but the site focuses on the player base as the whole. The database consists of 600k+ players and more than 6bn battles and 115M individual player-tank records. This is just a hobby project for me to learn R, MongoDB and data analysis. 

 

The Relative WR graphsare maybe the most interesting part at the moment for the general player population. It tells how how many percentages higher WR in average the players reach in that tank vs. their other same tier tanks. This is very similar to BlitzStars Relative WR except it compares to same tier tanks only and the comparison point does not include the tank itself. This is the 2nd best measure of tank performance I am aware of. The best one is the Delta WR I used in the old Google Sheets graphs, but I have not had time to implement it yet to the new code base.

Hint: The tanks leading the chart are your best picks for stats-padding

 

 

I have few things in in my backlog including:

• Detailed pages for individual tanks
• My "Delta WR" based tank performance analysis (if someone remembers those old charts)
• Graphs per Player skill-level
• Blog posts of analysis of different topics

 

But even now there are lot of graphs (1956 to be exact) to digest even for the most stats-hungry tank-nerds.  

 

All the feedback is welcome. You may reach out me here (I am more active on EU forums) or via Discord (Jylpah#4662) or email Jylpah@gmail.com

 

Yeah, Nerdy "stuff"

 

PS. You may be also interested in my Python scripts to upload and analyze replays

Edited by Jylpah, 26 March 2021 - 09:25 PM.

[acrisis]

Big o7 to you, sir!

Will be diving soon  

[acrisis]

Tons of great work and nice graphs.

You have covered a lot of ground! 

 

 

The one thing that stood out, in a sad way, is that everything is now centered around tier 8. 
 
Does your access allow information gathering about platform, devices, ... 
Any indication supremacy vs regular battles?

Or it is still simply battle and tank info based on personal stats.  

 

[Jylpah]

acrisis, on 16 May 2020 - 02:04 PM, said:

Tons of great work and nice graphs.

You have covered a lot of ground! 

 

 

The one thing that stood out, in a sad way, is that everything is now centered around tier 8. 
 
Does your access allow information gathering about platform, devices, ... 
Any indication supremacy vs regular battles?

Or it is still simply battle and tank info based on personal stats.  

Hello, 

 

WG API does not give access to stats by user device type, nor does it give stats for supremacy/encounter separately. I don’t there are usable stats for rankings either. WG’s Blitz API is even missing things like tracking/spotting damage while those are available for WoT Console. 

[acrisis]

I thought that would still be the case, but figured I'd ask. At times I'd love to slice into the player base and see what is what. 

 

This 

 

Really drives home how WG killed low and mid tiers, if I understand it correctly, that this is showing the tier battles that all players engaged in. 

 

And it seems consistent with prior months. Not just a fluke where people were trying to grind credits during an event. 

 

[_Tactician]

Interesting site

[Icefang50]

Can we get this thread pinned?

 

(edit)...asking for a friend

Edited by Icefang50, 16 May 2020 - 09:10 PM.

[Jylpah]

acrisis, on 16 May 2020 - 04:02 PM, said:

I thought that would still be the case, but figured I'd ask. At times I'd love to slice into the player base and see what is what. 

 

This 

Really drives home how WG killed low and mid tiers, if I understand it correctly, that this is showing the tier battles that all players engaged in. 

 

And it seems consistent with prior months. Not just a fluke where people were trying to grind credits during an event. 

 

 

Hello,

 

While the tiers played graph likely holds roughly true, there is one thing know that is not yet explained clearly at the site. Even the site has 600k+ players in the database, that is nowhere near the entirety of the Blitz player base and there is a selection bias:

 

It seems Blitz has huge number of players who play 1) just few games and quit, or 2) play just few games per update and leave the game after the next event/WG marketing /update/whatever. 10% of the players recorded in the stats play 8 battles or less during an update. And that is only of the players who are recorded in the database. And these players tend to be better and more active than a true average Blitz players due to the way how the site harvests account_ids - from BlitzStars (Thanks SockRobber) and WoTinspector.com replays (Thanks reven86). One ends up the BlitzStars only if someone (e.g. yourself) looks at your stats there. The same way replays are being uploaded by more active players who tend to play more higher tiers. The probability for someone playing 50 games at tier I-III during an update to get recorded in a replay submitted to WoTinspector.com us is lower than someone playing 500 battles at tier VIII.

 

Most clear evidence of this bias is the Player WR distribution chart that shows average WR above 50%. While it is theoretically possible for the average WR to be above 50% it would require that the players who play more have lower WR than those players who play less. And that is not what the stats are showing otherwise but vice versa. Therefore the dataset is missing large number or inactive and below average players.

 

Once I have time, I will experiment if I can construct a statistical model of the player base based on few assumptions (distribution of battles and "skill" / WR)... because ...err.. numbers are fun

Edited by Jylpah, 19 May 2020 - 07:40 AM.

[acrisis]

Thanks for the clarification.

 

So, yes, the selection is limited and hinges on proficient players who likely are playing tier viii mostly for economic reasons. 

[__V_O_P__]

acrisis, on 18 May 2020 - 10:56 AM, said:

Thanks for the clarification.

 

So, yes, the selection is limited and hinges on proficient players who likely are playing tier viii mostly for economic reasons. 

perhaps “...the selection is biased by prolific players who ...”

 

 

[__V_O_P__]

So here’s a question! Could we identify the impact of winning and losing? What I mean by that is could we use the data to take into account a predicted average damage accounting or adjusted for wins and losses?

Winning teams score more damage (mostly) and losing teams do less. Could we compare the two generally and factor that out?

[Jylpah]

__V_O_P__, on 19 May 2020 - 12:36 AM, said:

So here’s a question! Could we identify the impact of winning and losing? What I mean by that is could we use the data to take into account a predicted average damage accounting or adjusted for wins and losses?

Winning teams score more damage (mostly) and losing teams do less. Could we compare the two generally and factor that out?

 

Definitely!

 

That is on my task list. As most of people know, for a while one can do very well (or very poorly) in a tank and reach very low (high) WR since there are other factors (e.g. 13 other players) that impact to the outcome of a particular battle. Over the time the impact of the "other factors" will average out and it is only one's skill reflected in the WR. But that takes _lot_ of battles. As well explained here, it takes ~10k battles for one's WR to reflect your underlying skill level with +/-1% accuracy - assuming constant skill level of the player and the player base as the whole. If (read: when) the player skill level changes, even more battles are required to reach +/-1% accuracy (with 95% confidence level).

 

But with the stats it is possible to calculate estimator models for WR based on one's "battle output stats" (damage, kills, spots, etc). Unfortunately WG API does not give tracking/spotting damage for Blitz (it does for WoT Console), but the model can be rather accurate still. Once the time permits, I am planning to add separate pages for tier/vehicle class (comparison between the same tier meds/HTs/LTs/TDs along these lines.

 

Btw, the second plot shows well how the battle count impacts on the accuracy of the estimate (0.1 == 10%):

 

 

On those days all the analysis was ad hoc therefore very time-consuming, but now I have a publishing & template system in place so once the analysis is done once, the respective pages and graphs can be generated automatically. But before buffing the expectations too high, I have to say it will take a month before I get it done. Lot of things to do also out side Blitzanalysiz.com site (including actually playing the game). But the vehicle class / tank-specific analyses was the idea I had when I started building the site.

 

Cheers,

 

Jylpah

[Icefang50]

Bumping this so it doesn’t get lost behind various “locked” threads

[Jylpah]

Hello fellow stats-nerds,

 

I went to write a “post” about measuring tank performance and Relative WR.

 

https://blitzanalysiz.com/blog/2020-06-08_measuring_tank_performance/

 

Measuring tank performance

 

Over-Powered (OP) tanks are maybe the 2nd most popular topic on Blitz YouTube videos and online chats - right after “the Matchmaker”. I have long presented that if a tank is “OP”, it has to be visible in statistics. Otherwise there is only qualitative / subjective views left and those come in all sorts.

 

 

People are susceptible to all kinds of biases in their thinking. Getting ammoracked by a Death Star will raise suspicion that Death Star is an OP tank even it is actually the worst tier X tank. Anecdotal experiences distort opinions and no one remembers those countless of battles where a DS was at the bottom on the list. Therefore, tank performance is not a thing that we should vote about, but a thing we can measure with statistics.

 

How to measure performance in the game?

 

Let’s discuss first how to measure performance in the game. I am a proponent of win rate being the best measure for performance (player or tank)– not average damage, not average kills, and not speed, not alpha damage or any other attribute or characteristics. The reason for choosing win rate over other variables is the fact that winning is the objective of the game and all the damage, kills, spotting etc. are just means to win the game. Why to measure the proxy variables when you can measure the final variable itself?

 

There are caveats in using WR as performance measure:

 

1. It requires lot of battles for WR to settle near one’s performance level due to both MM and RNG: It take 400 battles to reach +/- 5% accuracy, 10000 battles to reach +/- 1% accuracy and 1 million battles! to reach 0.1% accuracy with 95% confidence level. Check this link at PC WoT forums for details.
2. Platoon rate impacts on WR but cannot be distilled well from the stats since WG does not publish detailed platoon rate per tank played. Platooning with a good player can lift one’s WR 5-15%.
3. Career WR measures historical average, not one’s present performance level, and reacts slowly once the player has lot (10k+) battles.
4. WG’s new “newbie MM queue” has distorted the Career stats for rerollers big time. This distorts both tank and player average WR. (Just ignore global & career WR).
5. Some tanks are more powerful than other. Comparing different players’ WR in different tanks or global WR does not tell much.
6. Different tiers have different level of difficulty. Global / Average WR is close to useless in measuring player / tank performance.
7. Stock tanks’ performance is significantly lower compared to maxed-out tanks’ performance.

 

But other performance measures have issues too and can be gamed; Easiest way to increase average damage play more high tiers and more TDs, WN8 can be gamed by playing popular tech three tanks that are difficult for below average players and not too popular among the unicums. Despite all the issues related to WR, I consider it the best performance measure over large number of battles and in case of tanks, over large number of players since it measures directly the objective of the game (=to win battles). It is also more understandable measure vs. somewhat abstract indexes. But I believe performance indicators like WN8 which based on “input stats” (average damage, kills, spots) give more accurate view on players’ short-term performance (< 100 battles). Now going back to the tank performance.

 

So Average WR it is then, right?

 

Not so fast. Average WR of a tank is a good starter, but it itself has its own biases. Let’s have a look on two tier IX mediums: AMX 30 1er prototype and Prototipo Standard B. Everything here is based on 6.9 data.

 

 

Tank	Average WR	Players
Prototipo Standard B	52.3	7622
AMX 30 1er prototype	56.5	3414

 

Both the tanks have been also played by thousands of players, but the AMX seems to have significantly higher average WR. Many would be tempted to claim the AMX is a better tank than Standard B. But is it?

 

Continue reading at blitzanalysiz.com

 

Edited by Jylpah, 09 June 2020 - 07:22 PM.

[_Crusader6_]

Fantastic read.

 

 

[Jylpah]

_Crusader6_, on 09 June 2020 - 08:08 PM, said:

Fantastic read.

 

Thanks!

 

It's bit on a long side. I started writing it, learned few things on the road, debugged my algos, ended up revising the site's Relative WR algo and filtering system and finally managed to finish the article.

 

Tonight I will actually play the game 

[_Crusader6_]

I found the new bit about the AMX 30 1er and the STD to be very interesting. 
  I do more dmg in the Std - but don’t win as much.  Mind you I haven’t played the 1er much and the std was solo while the Proto was about 1/2 tooned. 
  
I think it really gives thought to truly what OP means.  
 

 

[Jylpah]

_Crusader6_, on 10 June 2020 - 01:29 AM, said:

I found the new bit about the AMX 30 1er and the STD to be very interesting. 
  I do more dmg in the Std - but don’t win as much.  Mind you I haven’t played the 1er much and the std was solo while the Proto was about 1/2 tooned. 
  
I think it really gives thought to truly what OP means.  
 

 

 

It is good to remember that these are average over hundreds / thousands of players. There are always players who happen to do better in one tank over another or who have had a more successful run (MM) in one so far. Statistically it takes 400 battles to have one's WR to match their skill level with +/- 5% accuracy with WR (with 95% confidence level).

[Texas_Tyrant]

Great read as C6 said. I missed the conclusion though. Which one is better stat wise?

Texas_Tyrant

[_Crusader6_]

Texas_Tyrant, on 10 June 2020 - 07:03 AM, said:

Great read as C6 said. I missed the conclusion though. Which one is better stat wise?

Texas_Tyrant

The answer is what tank you personally can play better...