Gentleman Jon —ETF2L Season 17 regular season has drawn to a close and we now face the blood and guts, do or die, go-home-with-your-shield-or-on-it playoffs to decide who will bathe in the rarefied glow of ultimate victory. I’ll be doing a little preview of the matches as they come along, starting with Dotter vs Awsomniacs. After that for the terminally short of anything better to do there’s some ramblings describing how I calculate the rating.
Before we begin, no logs involving the dropped team now known as Bonne Ambiance 2k14 have been used.
Awsomniacs vs Dotter
There’s very little to separate Dotter and Awsomniacs during the regular season, the Frenchies only finishing one point ahead in the final standings. Their individual match went the way of Awsomniacs but only by the narrowest of margins via a golden cap win on Process and a 3-2 victory on Viaduct. It seems from their league records that Awsomniacs have a miniscule edge.
In a season littered with default matches and drop-nullified results this can be misleading. Examining their records and applying a technique invented by Bill James (a baseball stats guy) called Pythagorean Expectation on their cumulative round scores for each regular season match and their expected league points should be 23 for Dotter and 25 for Awsomniacs. Both teams fell a little short of their just rewards, but the gap looks slightly more favourable for Awsomniacs than their league standings suggest.
In addition to this Dotter did play a number of games with their reserve Larsa, although statistically he was very similar to Knutsson.
Demomen
For newbies or those with a short memory, the yellow bar is the new improved rating and the blue bar is the player’s DPM (both measured by the bottom axis) and the red bar is the player’s KD, measured by the top axis. A high KD with a comparatively low DPM might indicate passivity, while a low KD with a high DPM aggressive play.
As you’d expect Ryb playing the more traditional defensive demo role is statistically superior to Exon whose roaming style get him killed a lot more frequently. The difference is mostly in KD, with the damage output fairly similar. Exon does a lot of dirty work for the team, making space and appearing as an unexpected kamikaze presence which is bad news for him, but good news for his team mates who don’t have to do the same.
This illustrates the relationship between damage and kills for the individual players. The yellow bar is the average damage they get for every landed shot and is measured against the bottom axis. The blue bar is the amount of damage they had to cause to register a kill or assist and is also measured against the bottom axis. The red bar is the number of kills achieved per minute, which is measured by the top axis.
Ryb kills slightly faster and has to put out slightly less damage for each kill/assist he gets. As you’d expect when roaming Exon is isolated more making it harder to register assists so comparatively is behind in both these departments, although not to an enormous degree. He does have an edge in the average damage for each shot he lands though, as he spends more of his time right in his opponent’s faces.
For the present it looks like a narrow advantage to Dotter on the demoman front but with compensation for Awsomniacs.
Soldiers
The two pocket soldiers seem to be pretty evenly matched in this overview with Tek getting a small edge against Zebbosai. Both play a pretty similar style taking full ubers as well as a lot of heals and expecting to spearhead their team’s pushes.
T-Mac has a clearer edge over Knutsson in both KD and damage, and examining his kill record in more detail shows he’s much more likely to pick a medic.
There are some interesting differences between the pocket soldiers, despite doing similar damage with their shots, Tek has to cause far more damage to register a kill or assist. As a result of this (and as you might expect as we’ve established they have similar damage output) Zebbosai gets a faster kill rate. To me this says that timing is more of a factor for Zebbosai’s aggression and they do a better job of engaging when they can get kills, perhaps more heavily basing their play around the uber or popping more riskily when fully engaged.
T-Mac continues the trend of having a clear statistical margin over Knutsson who has to put out more damage to get a kill or assist and gets them more slowly. Both roamers get higher average landed shot damage than their pockets because they’re bombing. The one thing in Knutsson’s favour is that T-Mac seem to see his own Medic quite a bit more (partly due to Exon’s roaming exploits), which could be contributing to some of these differences.
Scouts
Flippy, more normally known for his sniper exploits, is having one of his better seasons on scout coming in narrowly ahead of wltrs who is rediscovering some of his old mojo. Flippy is playing a slightly more conservative game than his opposite number staying alive marginally more often while wltrs has slight edge in damage.
Beavern is a little way behind on the lowest heals of all the scouts, while Dmoule with the most heals seems to be amongst Bulle’s protectors, a state of affairs you’d expect to lead to a lower DPM but the low KD is disappointing.
The two French scouts both spam more damage getting lower damage per landed shot and doing more damage for every kill/assist they get, and they both achieve a lower kill rate because of it. Beavern on the other hand keeps his powder dry until he’s landing the meat shots with the highest kills per minute, the most damage per landed shot and the lowest damage required per kill or assist.
Wltrs is a little more circumspect but outside the broad categories the Swedes look like the more efficient killing machines while the Frenchies look to clear up like vultures on Tek’s trail of destruction. Looking at their kill records both Swedish scouts are more likely to get to the enemy medic.
As a pair the French scouts get more healing and once again this is probably a result of Exon’s roaming freeing up some of the medic’s time. Even so it seems like the Swedes have a narrow edge here.
Sniping
One the biggest game changing switches any team can make is a scout switching to sniper and going on a headshot spree, and a game featuring Flippy and two Swedish scouts with a reputation for sniping demands analysis.
The blue bar is kills per minute, measured against the bottom axis and the yellow bar is headshots per minute, also measured on the bottom axis. Red is uber drops (just the simple total) measured on the top axis.
There’s a contrast of styles between Beavern and Flippy, with Beavern concentrating on the kill via any means necessary and Flippy playing for the purist pleasing headshots. Wltrs stats mirror Flippy’s but to a lesser degree. Interestingly wltrs has a lot more time on sniper than Beavern, but with less success. Purely from a statistical standpoint it looks like Beavern should shoulder a larger share of the sniping burden.
Medics
In this graph the blue bar shows the kills and assists per death and is measured against the bottom axis, the red shows healing per minute and is measured against the top axis and the yellow shows assists per minute also measured against the top axis.
When examining the medic stats we’re partly looking at the team stats as their survival is so dependent on their teammates. With this in mind it appears that 2nuts is more combat focused aggressively following his team into danger and racking up the assists whereas Bulle stays a bit safer, buffing his team rather than healing the fragging player quite as often, but it’s also worth noting that these stats reflect some of the differences between the pocket soldiers they spend most of their time healing as well.
Their uber per minute rate is very even, but the drop rate significantly favours Bulle, which would reflect 2nuts more obvious proximity to danger although a slight building advantage given that he has extra ubers to drop. Given how important ubers can be to the results of the game and how closely matched these two teams are it’s the kind of difference that can be the difference between winning and losing.
Ultimately there’s no clear advantage here, it’s a matter of choosing your medic style to fit the team. Both medics must support a dominant aggressive pocket soldier but 2nuts also has Ryb in relatively close proximity whereas Bulle will have opportunities to spread the healing around a little more.
A Prediction, sort of
This is a tough one, the stats suggest that Awsomniacs have tiny advantages with the stronger pocket soldier and strongest scout and that Dotter have the weakest roamer and are more likely to drop uber. On the other hand the Swedes should be able to get more out of their demo and the Frenchies have the weakest scout while theirs looks like assassins.
Their pocket styles are similar and the roaming soldiers are fairly traditional but beyond that there are significant differences, the roaming demo versus the relatively passive, the ambushing meat-shotting close range scout play versus the damage spamming scavengers and the varying risk taking of the two medics.
This could easily come down to a favourable map pick (I believe Awsomniacs have the first pick), or an external issue like the famously unreliable Swedish internet. If there’s any justice there should be golden caps aplenty but if forced to make a choice I would go Awsomniacs because they’ve managed to capitalise on their narrow advantages against Dotter already this season, although it’s hard to have confidence in this pick.
So this rating thing, what is it?
I implemented this thinking that my ramblings on the subject had been posted previously but looking back I realised I’d only made some vague references to it, so this is the real deal.
Damage and kill ratios are the most common ways of looking at players performances, but frequently you find a player has a good score in one but a low score in the other and it’s not clear which one matters more. To rank them we have to combine the scores somehow, and to do that we need to know the weight to give them.
I looked at 6v6 results and how the differences in various metrics between the teams related to the final result. The findings for damage and kills are illustrated below:
The interesting bit (you’ll have to trust me on this) are the R2 numbers which show the strength of the correlation of the stat to the final score, with 0 being no correlation and 1 (or -1) being a perfect correlation. Putting out a lot of damage is important, but kills are more important. Of course the two are related and in any given team some players will be required to do one or the other more, but this gives some indicator of the relative importance of the two.
One of the counter-intuitive things about damage is that the more one team puts out, the more likely it is that their opponents will score high damage too. Obviously kills don’t work this way, if you die repeatedly you’re not more likely to jump up kill your opponent, although you will have probably done him some damage. This doesn’t help much while he caps the point.
So to reach the rating I take the two correlations and roughly scale the two values to them and combine them to reach a rating. (DPM * 0.6) * (KD * 0.8). Mathematically dubious? Almost certainly, but I won’t be submitting it to the Nobel committee this year.
This doesn’t overtly recognise the tactical necessities of the game, but I think it does set an expectation for a level of damage that has to be done if a player is going to be aggressive and risk their life more often.
Improving the rating
We all know that not every kill is equal and recently I’ve noticed players who pick up easy frags benefitting unduly from this formula as they protect their KD without facing the really hard targets (obviously not to get a good rating, no one really cares, but you get the idea).
So I looked at the individual classes and how their deaths related to the final results.
Please excuse some of the graphs slanting in different directions, it doesn’t affect the results, what matters is the R2 value. Straight away it’s obvious there’s something wrong. Everyone knows that soldiers are worthless and their lives should be thrown away. So why are they rated higher than demos? The simple answer is that they have a stronger correlation because there’s two of them on every team, and their deaths are therefore a better barometer of how the match went than any other classes.
The same goes for scouts, although their individual KD seems to have less importance. Finally statistical proof of what everyone always knew, the swines take care of their own skin above all else including their team.
The easy way to resolve it would be to say, half a pair must be half as important and just take half the number, but that would be a bit lazy and isn’t really convincing logically. So, I took a look at how kills on individual soldiers and scouts rather than as a pair in a given match affect things
Separating the individual performances out produces even less correlation than halving the previous results and produces a switch between the importance of the soldier and scout frags. This is probably down to the roamer’s tendency to throw his life away regularly. This also gives us a new problem, surely a demoman isn’t worth 3 soldiers? A demo and medic combo are not going to be able to fight off 6 soldiers, which is what these stats are telling us the pair of them are worth.
The reason is that the demo to some extent and medic’s deaths often occur after the rest of the team is dead, so the team’s deaths are being counted twice by proxy. Instead of solving problems, they just keep mounting up. On the one hand we have the rest of the team’s deaths being counted too many times by implication and on the other over representation of the importance of certain classes because of their number.
I decided to cut this Gordian knot by simply taking the halfway point between the two soldier values and the two scout values as this produces something plausible and (superficially at least) statistically defensible, while acknowledging that their presence in numbers contributes to the survival of their team mates. A better method might be to look at the average team compositions when those classes die to filter out the deaths of other players as a factor, I may look at that in the future so this represents a best fit for now.
This leaves us with the final values of 0.6 for demo and medic, 0.4 for soldier and 0.33 for scouts which I’ve used be using to refine the KD used in the ratings. This means that demo ratings will generally be lower as they will generally get lower rated frags, and scout will enjoy a ratings boost every time they kill anyone who isn’t a scout.
The final conundrum is the fluctuating value of the medic, rising slowly until uber is achieved and sharply dropping to near zero when it runs out. That’s not something I’ll be solving here, but it could be a focus for future analysis.
Normal
0
false
false
false
false
EN-GB
X-NONE
X-NONE
Originally published on VanillaTV (vanillatf2.org) on 2014-03-22 by diemasher.
Recovered from the Internet Archive (archived copy) as part of tf2esports.com's preservation of VanillaTV's TF2 coverage. The original site is no longer online. All rights remain with the original author; reposted for historical reference.