183 replies to this topic

[jylpah]

Hy, 

 

Being always curious about Blitz statistics and vehicle balance, I decided to crunch Tier X data kindly provided by BlitzStars.com admin, Commander SockRobber. The data sets (one from end of January 2017 and another from end of August) have stats of 40k players from all the four servers (10k per server). I have calculated deltas between the two data sets to be able to see how the 40k randomly selected players have performed during this February-August period. All the analysis below is based on this February-August data, not on historical cumulative data. The analysis is done using free R Studio. 

 

(Simple) Weighted Linear model

 

First I built a weighted linear model to estimate WR based on player's "battle output stats" (average damage, kills,   spot rate, etc.). The model does not consider different tanks, classes etc. Since different players have different number of battles in a tank during the period, I have used "weights" feature in the R's glm() function to give indicate that the stats are averages of number of datapoints (i.e. all.battles.diff). I noticed also I get better results if I use only those datapoints where a player has at least 100 battles in a tank during the period for calibrating the model.

 

Spoiler

 ## Weighted General Linear Model (Avg dmg * Kills) SIMPLE ------------------------------------------------------------------------
tier_10.wglm2 <- glm(WR.diff ~  avg.dmg_dealt.diff  + avg.dmg_received.diff + avg.frags.diff + avg.frags.diff*avg.dmg_dealt.diff + spot_rate.diff + hit_rate.diff + avg.cap_pts.diff + avg.cap_defense_pts.diff, weights = all.battles.diff, data=dat.diff.tier_10[all.battles.diff > 100,])

 

 

About the model's accuracy / WR estimate error

 

The model estimates the WR with 5% accuracy for most of the players with more than 50 battles in a tank. While this may sound a poor accuracy, this is simply due to number of battles players have in a tank. 

 

 

 

When the WR estimate error is plotted against number of battles in tank, we can see how Randomness (i.e. "the teams" )  can cause large estimate errors with low battle count. To reach 5% accuracy one needs ~500 battles in a tank. Just remember that when you next time rage about losing 0.01% of your WR "due to teams". One dot in the graph below is represents one player in one tank. To reduce noise/randomness, I have selected only data points with more than 100 battles in a tank. 

 

 

 

What does the model tell in overall? 

 

Spoiler

 > summary(tier_10.wglm2)
Call:
glm(formula = WR.diff ~ avg.dmg_dealt.diff + avg.dmg_received.diff +
    avg.frags.diff + avg.frags.diff * avg.dmg_dealt.diff + spot_rate.diff +
    hit_rate.diff + avg.cap_pts.diff + avg.cap_defense_pts.diff,
    data = dat.diff.tier_10[all.battles.diff > 100, ], weights = all.battles.diff)
Deviance Residuals:
    Min       1Q   Median       3Q      Max  
-6.2605  -0.3658  -0.0037   0.3556   5.7453  
Coefficients:
                                    Estimate Std. Error t value Pr(>|t|)    
(Intercept)                        2.120e-01  6.698e-03  31.647  < 2e-16 ***
avg.dmg_dealt.diff                 8.481e-05  2.040e-06  41.568  < 2e-16 ***
avg.dmg_received.diff              3.648e-06  1.940e-06   1.881     0.06 .  
avg.frags.diff                     6.253e-02  4.235e-03  14.765  < 2e-16 ***
spot_rate.diff                     3.576e-02  6.154e-04  58.111  < 2e-16 ***
hit_rate.diff                      3.653e-02  7.372e-03   4.954 7.32e-07 ***
avg.cap_pts.diff                   1.875e-02  1.079e-03  17.369  < 2e-16 ***
avg.cap_defense_pts.diff          -6.355e-03  8.690e-04  -7.313 2.73e-13 ***
avg.dmg_dealt.diff:avg.frags.diff  8.423e-06  1.656e-06   5.087 3.67e-07 ***
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
(Dispersion parameter for gaussian family taken to be 0.306949)
    Null deviance: 25891.2  on 17557  degrees of freedom
Residual deviance:  5386.6  on 17549  degrees of freedom
AIC: -65951
Number of Fisher Scoring iterations: 2
> tier_10.wglm2.R <- 1-5386.6/25891.2        # Pseudo-R
> tier_10.wglm2.R
[1] 0.7919525

 

While the above R output can be cryptic to read, the model provides the following info:

• The simple model can explain 79.2% (tier_10.wglm2.R) of the variance which is decent result. The rest is "randomness" and "unknown factors". 
• Intercept "variable" i.e. the constant is 21.2%: Therefore,  the model estimates a full-time-AFK player would get 21.2% WR at tier X. I did the same analysis for tier VI and there a full-time-AFK would get 33.4% WR. 
• Average Damage (avg.dmg_dealt.diff): Increase of 1000hp will increase your WR by 8.5%
• Kill sper Battle (avg.frags.diff): One kill/battle will increase your WR by 6.2%
• Spot rate (spot_rate.diff): One spot per battle will increase your WR by 3.6%
• Hit Rate (hit_rate.diff): 10% (nominal) increase of will increase your WR by 0.4%
• Average Damage Received (avg.dmg_received.diff): is not statistically valid (p-value 0.06)

 

Average damage vs WR

 

There is a strong linear correlation between Average Damage and WR. Please note that this graph bundles all tank types together. Still, doing damage alone is not enough. For example with 2000 average damage, one can have a WR between 40-60% depending other output factors like average kills, spot rate, other factors not captured by WG data (e.g. Assisted Damage, Blocked Damage, tactical mastery resulting wins, but not measurable with data) and pure luck (aka "the teams"    ) . As said, the element of luck decreases as the number of battles increases. R-value is 0.71 indicating clear correlation. 

 

 

Kills per battle vs WR

 

There is also clear linear correlation between average kills per battle and WR. Again, just kills won't win a battle, but having 1 kill per battle on average at tier X can result WR between 40% and 70%. Some of the most extreme data points might be errors in the WG data (yes, there are several people having simply incorrect data like average damage higher than the total HP pool of the reds...). The R-value is 0.68 indicating clear correlation. 

 

 

WR vs Average Damage + Kills per Battle

 

Here is a plot of WR against average values of Average Damage and Kills.

 

 

WR vs Average Damage + Kills per Battle [Heavy Tanks]

 

 

WR vs Average Damage + Kills per Battle [Medium Tanks]

 

 

WR vs Average Damage + Kills per Battle [Tank Destroyers]

 

 

 

 

Spot rate vs. WR

 

This is where it gets interesting. R-value is only 0.09 meaning that that there is only a weak correlation between WR and Spot Rate. You can spot by YOLOing hard and by doing it right. 

 

 

WR vs Survival Rate and Spot Rate 

 

The art of spotting is to spot without dying. Color in the graphs shows the player's WR. Couple of interesting findings can be made: 

• The very best do NOT spot the most, but they spot "reasonable amount" ~1.5-2.0 spots per battle.
• Camping does not win games: High Survival Rate combined with low spot rate is a recipe for failure. 
• In general, better players spot more, but this plateaus ~60% WR. Players better than this actually post less and are likely more cautious in their game play.

 

 

However, the graph above bundles all tank types together.  Let's see how does the data look per different tanks class. 

 

WR vs Average Cap Points vs. Survival Rate and Spot Rate PER TANK TYPE

 

Heavy Tanks

 

From the data it is clear that:

• Players who spot more in a heavy tank (i.e. play it more aggressively) tend to win more.
• Max spot rate in heavy increases up to 60%, but then starts to decline. 
• The very best spot ~1.2-1.7 tanks per battle
• Camping in a heavy tank does not work - what a surprise ...  

 

 

Medium tanks

 

• First of all, even the very best Tier X drivers do not reach such a WRs in medium tanks as they do in heavy tanks. Let's remember that the time period is Feb-August 2017 and the Prammo Nerf was introduced in 3.9 (June 2017)
• The best Medium tanks drivers spot 2 tanks per game on average.
• What comes to "potatoes", camping is better than YOLOing... (up to 45%...)

 

 

Tank Destroyers

• Players reach higher WRs in TDs than in Medium tanks
• Very few players are doing super-bad (sub 35%) in Tier X TDs. If one gets only one shot per battle having a big gun helps 
• Spotting in TD is not the way to go. The very best spot ~0.8 tanks per battle. 

 

 

Capping points show no correlation with WR. The R-value is only 0.01 meaning Average Capping Points do not explain someone's WR pretty much at all. 

 

 

Average Cap Defense Points vs WR

 

There is very weak correlation between Average Cap Defense Points and WR (R-value is 0.11). However, the data says it is more important to defend the cap than to actually cap (R-value 0.11 vs 0.01).

 

 

 

 

 

Edited by jylpah, 07 December 2017 - 08:15 AM.

[minitelrose]

Deep work. As always.

my respects, commander Jylpah ! You are clearly working for the cause.

[Huwie_]

My eyes

[minitelrose]

Huwie_, on 25 September 2017 - 03:06 PM, said:

My eyes

 

it's ok. Just make sure to not press to open the spoiler tabs. That's the pain.

[jonty_2014]

Impressive number crunching - keep it up.

[LoadAimFireRun]

OMFG, please help; I lost 0.01 percent of my winrate due to teams !!

[Huwie_]

minitelrose, on 25 September 2017 - 02:16 PM, said:

 

it's ok. Just make sure to not press to open the spoiler tabs. That's the pain.

 

that was my mistake

[Jukkis74]

The ’intercept’ difference between tiers 6 and 10 is rather large, which is very encouraging. If that withstands a closer scrutiny, clueless players indeed will get punished time and time again in highest tiers. (As they should be.) 

[minitelrose]

Huwie_, on 25 September 2017 - 05:04 PM, said:

 

that was my mistake

 

i know, right ? I had to open to know it too...

[jylpah]

the content moved to the first post.

Edited by jylpah, 27 September 2017 - 10:30 AM.

[jylpah]

smash_LGN, on 26 September 2017 - 02:54 PM, said:

It works as intended. 3500 average damage, 2 kills/battle, 43% WR. 

 

There is clearly something wrong with his data for this period. He is not the only one. I have cleared many of those away already. 

 

R-code

 > dat.diff.tier_10[(account_id == 2004807932) & (tank_id == 12305), .(account_id, tank_id, name, all.battles.diff, WR.diff, avg.dmg_dealt.diff, avg.frags.diff, spot_rate.diff, hit_rate.diff, survival_ratio.diff)]
   account_id tank_id         name all.battles.diff  WR.diff avg.dmg_dealt.diff avg.frags.diff spot_rate.diff hit_rate.diff survival_ratio.diff
1: 2004807932   12305 E 50 Ausf. M              314 0.433121           3470.016       1.996815       1.792994     0.7627876           0.7707006

 

https://www.blitzstars.com/player/asia/playerwot1406/tank/12305

 

 

 

Edited by jylpah, 26 September 2017 - 07:59 AM.

[jylpah]

Added more data to the first post.

 

Teaser

 

 

 

Edited by jylpah, 27 September 2017 - 08:26 AM.

[minitelrose]

The spot/survival graphs are really  a fantastic addition. The entire article is just rock solid work !

my first reaction however is that the winrate scale is not unified across the maps for tds, heavies and meds. I find it disturbing to try to estimate the shape and location of blue, but blue carries a different meaning in the three graphs. In other words if I want to see what defines me, for exemple 55-60%,  that "me" becomes a different color in each graph. 

[jylpah]

minitelrose, on 27 September 2017 - 06:43 PM, said:

The spot/survival graphs are really  a fantastic addition. The entire article is just rock solid work !

my first reaction however is that the winrate scale is not unified across the maps for tds, heavies and meds. I find it disturbing to try to estimate the shape and location of blue, but blue carries a different meaning in the three graphs. In other words if I want to see what defines me, for exemple 55-60%,  that "me" becomes a different color in each graph. 

 

Bugger, I’ll fix that 

 

EDIT. Fixed 2017-09-28 

Edited by jylpah, 28 September 2017 - 12:07 PM.

[Cancracker]

And the key to the matter is and I quote " pure luck (aka "the teams"    ) "

Of course some get always more "luck" and are at 70% wr while others get always less and are at 40% wr. And we all know how this "luck" is implemented...

[minitelrose]

Cancracker, on 27 September 2017 - 04:51 PM, said:

And the key to the matter is and I quote " pure luck (aka "the teams"    ) "

Of course some get always more "luck" and are at 70% wr while others get always less and are at 40% wr. And we all know how this "luck" is implemented...

"Luck"=\="noobness"

[Sir_Bobby]

Cancracker, on 27 September 2017 - 03:51 PM, said:

And the key to the matter is and I quote " pure luck (aka "the teams"    ) "

Of course some get always more "luck" and are at 70% wr while others get always less and are at 40% wr. And we all know how this "luck" is implemented...

 

but your tier X dmg output, spotting, kills and survival would put you at about 50% on jylpahs charts, which is exactly where you are.

 

the whole thing is very similar to wotbstars expected winrate, in which you actually seem to win slightly more than expected.

 

so for you at least it doesn’t seem that luck, good or bad, plays much of a part at all.

[Tijgerhaai_XIV]

This is just awesome stuff, jylpah.

[Jukkis74]

Best thread of the month material. And can grow to best thread of the year stuff.

[jylpah]

Cancracker, on 27 September 2017 - 11:51 PM, said:

And the key to the matter is and I quote " pure luck (aka "the teams"    ) "

Of course some get always more "luck" and are at 70% wr while others get always less and are at 40% wr. And we all know how this "luck" is implemented...

 

Had you read the whole post, you had also learnt how the randomness (“luck” ) - element disappears as the number of games increases. At 1000 games (in a tank) it’s all about skill, nothing about “luck”.