Heya, Go4ino here again with another SGC data analytics report. The original inspiration for this report was RebelFox on twitter asking if I could investigate if differences in performances existed between different champions with different types of player bases.
For example: Is there a difference between Ornn and Sona? Ornn has been played extensively by practically every top laner in SGC, while Sona is more or less exclusively played by Dean. Does this mean Sona has higher average performance? Or is it overall better to stick primarily to the meta?
Thankfully @pookarGG’s SGC data set exists and has concrete data to analyze. I did heavy data modification / rearranging / tinkering / etc to get the data sets I used in R-Studio.
As mentioned previously, the original source for this data was from Pookar’s SGC stats doc. In specific this report pulled data from the Champion By Player tab. I put all this data together into an Excel spreadsheet to read into R-Studio.
Pookar’s data takes match results data from almost every AM league up to 7/27, with only BIG League missing. Even with some matches being absent, there are 493 games present to analyze.
The sheet is heavily automated, meaning every tab of interest is updated when match histories are entered. I chose the Champion By Player tab as it allowed me to separate champions by position more easily to distinguish flex picks, and I felt like lumping flex champs together would skew results (eg: Top, Jungle, and Support Sett have vastly different statistics).
fig 1: Interactive dot-plot of player champ play rate vs win rate
note: only champions with at least 10 games played, and players with 10 games on a champ are shown.
To interact with the graph hover over the graph to display options, hover over data points to display critical information, and click on the positions in the legend to show/hide data points for a specific role.
The vertical lines at 20% and 60% of total play rate is where I drew lines for separating the 3 factor groups. I decided on these 3 groupings based off what I perceived to be 3 distinct groups of data, and picked 20% and 60% since they would divide the 3 groups fairly well. Of course, these separations are by no means written in stone and just my personal interpretation which I based purely off of Proportion of Picks, and my eyeball’s guessing power. Important to note is that due to the minimum game requirements this graph is naturally biased towards players who have played the most games. For example 100T FallenBandit moved to starting top during UPL playoffs and has only 8 games in the recorded data, meaning it is impossible for him to even be included anywhere on fig 1. Likewise, teams who play more games per set on average are more favored to have their players appear on the graph. Lets say Team X is extremely dominant, and 2-0s every opponent, but Team Y is less consistent and all their series go to game 3. This means that Team Y plays 50% more games than Team X despite playing the same number of Bo3 series, and inherently is biased in favor of Team Y.
Points to the left are more widely played meta champs, whereas points to the right tend to have players who represent a significant portion of a given champ’s playerbase. As such I have categorized the 3 playerbase groups as: Broad, Medium, and Narrow. You can think of Broad being meta champions, Medium being somewhat meta champions that certain players/teams may more heavily favor than others, and lastly Narrow is champions that are almost exclusively played by single players.
As expected, the further along the x-axis you go the fewer data points there are. But are there any differences to be noted?
So we now have a concrete definition for the 3 primary types of playerbases. This means we can look at each of the 3 groups and try and discern any trends, differences, anomalies, etc.
| Playerbase descriptive stats from fig 1 | ||||
|---|---|---|---|---|
| Playerbase | N | Distinct Champions | Mean Win Rate | Mean Champion Total Games |
| Broad | 74 | 23 | 54.4% | 141.39 |
| Medium | 10 | 10 | 56.74% | 52.70 |
| Narrow | 4 | 4 | 75.06% | 19.75 |
table 1
The Broad category has astronomically more data points than the other 2 groups by a mile, however there is plenty of overlap with a relatively meager 23 distinct champions. For example: Aphelios appears 9 different times in the Broad category (click the legend to change which positions show on fig 1 to make it easier to visualize). On the flip side, Medium and Narrow both are entirely made up of unique champion picks. It should be noted that it is impossible for any duplicate champions in the same role to appear in Narrow due to the requirements. Furthermore, while it is possible for duplicates to appear in Medium, it’s unlikely since all the champions in it are played somewhat widely and it just happens some teams / players value them higher. As for Narrow’s champions, 3/4 of them were explicitly named as examples when the topic was brought up with Hunter’s Nunu being a surprise point to me at least. This is one frontier where I absolutely hate the missing data from certain series / leagues / etc. because I’m sure there would be more points in Narrow I feel.
Now for win rates. I should preface this by saying to take this all with a grain of salt. As noted previously, the population sizes for Medium and Narrow are really small due to my limitations for games played. Even with those caveats I do still think it is worthwhile to examine the results to look for trends. With that being said, there is no significant difference in win rate between Broad and Medium champions aside from a slight bump upwards. I personally expected there to be at least some more of a difference in win rate. All the Medium champions do have decent sized playerbases though and are seen enough most players are at least familiar with the matchups, so maybe familiarity is the reason for only a slight deviation? I’m not entirely sure as it could be a variety of factors, and this is something I feel maybe the players themselves might be able to give good insight into this. Now for Narrow we see a massive jump to a win rate over 75%. Of course with only 4 data points there is a looming threat of outlier error, but at a glance it seems like there isn’t any massive deviations in the win rate as the median win rate for Narrow is 75.1% which is barely off the mean. With that in mind, I’m fairly confident in saying that having a player who pilot champions nearly no one else plays can yield some great results for their teams. A 75% win rate is insane, and keep in mind in most instances the champions in this category typically receive few changes (usually buffs) over entire seasons sometimes. This means that the potency of these picks remains the same over competitive splits, forcing enemy teams to either prepare for a pick that only really one person may play, or in most cases use up a ban. “Just use 1 ban no biggie” may be how some interpret this, but I want to paint a picture here to show this in action:
So it’s no secret that for the majority of the split, red side was forced to ban Varus in first ban phase every game. Now lets say your team is red side against SuperNova, you ban Varus of course, and then Always Plan Ahea’s A-sol. But now you only have 1 ban left in first phase, meaning a strong pick you wanted to ban slips through and SuperNova grab that champion. And the cherry on top is that Always Plan Ahea just locks in Syndra and still performs well.
That may sound like hyperbole, but it was fairly common Always Plan Ahea said himself.
“It for sure made atleast the first half of pick/ban easier with them banning asol every game. It would make it so that we could draft around getting certain picks because we only know that they are working with at most 2 bans”
A glance at the Team Win/Loss % - Side tab on Pookar’s sheet shows that the average blue side win rate is 53.75%, but SuperNova has a significantly higher average blue side win rate at 72.41% and having the enemy usually use a ban on A-Sol may be partially why.
Lastly, we have the average number of total games a champion has been played across all of SGC depending on which playerbase category the data point falls into. Unsurprisingly this number drops off heavily like in the column N. Turns out if you have 10-15 games on a champion but account for only 10% of the picks overall, then there’s going to be a very high total game count for that champion! For the Narrow category I do wish I had ban data to see what the overall (pick+ban)/total games ratio is for those points, which would make it easier to draw definitive conclusions. For Always Plan Ahea’s A-Sol he was able to share with me that SuperNova would always pick A-Sol if it managed to get through bans, but I lack the critical information for the other 3. Like are these players picking these champions in Narrow whenever they can? Or do they save them for specific situations?
This was a question that popped into my head for the 4 players in the Narrow category. While boasting high performance on those champions they’re known for is great, can they still perform when not on said champions? Because if they struggle on different champions, well then said power picks could be a potential weakness for their teams rather than a strength.
The column of most interest is Ratio which is just \(\frac{Column-2}{Column-3}\), as that can show if there is any statistically significant differences. A score of 1.00 means a player overall performs the same even if they aren’t playing their signature champion. Scores above/below 1 mean they perform worse/better overall when not playing their signature champion1. Ideally we want this ratio to be as close to 1 as possible, values significantly above 1 mean they struggle on different champions, and values below 1 are kinda pepega.
| Always Plan Ahea Averages | ||||
|---|---|---|---|---|
| Metric | A-Sol | Non A-Sol Champs | Difference | Ratio |
| Games | 15 | 45 | ||
| Win rate | 0.8 | 0.3855 | 0.4145 | 2.0752 |
| KDA | 5.77 | 5.18 | 0.59 | 1.1139 |
| KPAR | 0.6285 | 0.7163 | -0.0878 | 0.8775 |
| DPM | 656.15 | 604.29 | 51.86 | 1.0858 |
| CS/Minute1 | 8.8 | 7.672 | 1.128 | 1.147 |
| CS @ 101 | 80.73 | 80.226 | 0.504 | 1.0063 |
| Vision Score/Minute | 1.09 | 1.025 | 0.065 | 1.0634 |
| Enemy Jungle Camp CS2 | 1.87 | 2.906 | -1.036 | 0.6435 |
| Lane Diff @ 10 | -165 | 58.894 | -223.894 | -2.8016 |
|
1
Lane minions only, jungle camps are not included.
2
Number of enemy camps taken.
|
||||
table 2
Always Plan Ahea has ratios mostly close to 1 across the board, of note though is Win rate. Because while the other 3 all have significantly higher win rates on their respective champions, Always Plan Ahea has a ratio over a whole 2.0. Almost all of Always Plan Ahea’s other metrics are fairly close with Lane Diff @ 10 being the other significantly different metric2\(^,\)3, I wonder if that has to do with A-Sol’s tendency to roam a lot? I expected a significant difference in win rate, but not this much. It’s not like there are major changes in the other metrics, so it could be some other factor outside of the data set, but as it stands this insane dip in win rate can not be ignored.
| Dean Averages | ||||
|---|---|---|---|---|
| Metric | Sona | Non Sona Champs | Difference | Ratio |
| Games | 26 | 33 | ||
| Win rate | 0.7692 | 0.4322 | 0.337 | 1.7795 |
| KDA | 5.58 | 4.2936 | 1.2864 | 1.2996 |
| KPAR | 0.6723 | 0.6359 | 0.0364 | 1.0572 |
| DPM | 205.18 | 176.9464 | 28.2336 | 1.1596 |
| CS/Minute1 | 1.29 | 1.1173 | 0.1727 | 1.1546 |
| CS @ 101 | 14.15 | 11.4582 | 2.6918 | 1.2349 |
| Vision Score/Minute | 1.85 | 1.9236 | -0.0736 | 0.9617 |
| Enemy Jungle Camp CS2 | 0.31 | 0.0909 | 0.2191 | 3.41 |
| Lane Diff @ 10 | 408.54 | 106.4127 | 302.1273 | 3.8392 |
|
1
Lane minions only, jungle camps are not included.
2
Number of enemy camps taken.
|
||||
table 3
Like Always Plan Ahea, Dean also has a massive differential in Win rate albeit not as large of a gap. Lane Diff @ 10 is the other main massive difference4, with almost a 400% differential which can be interpreted to potential mean several things:
Any of the listed reasons could be true, but I do wish I had been able to interview FrostFire’s bot lane to get their perspectives. I can say for certain though Dean is very comfortable on Sona with almost as many games as Sona as games off Sona.
Likewise, all of Dean’s metrics (save for Vision Score/Minute) are above 1, with many being significantly higher5. I think it’s safe to say that Dean does perform notably better on Sona than when he is off Sona.
| Hunter Averages | ||||
|---|---|---|---|---|
| Metric | Nunu | Non Nunu Champs | Difference | Ratio |
| Games | 10 | 60 | ||
| Win rate | 0.7 | 0.5035 | 0.1965 | 1.3903 |
| KDA | 6.03 | 4.3153 | 1.7147 | 1.3973 |
| KPAR | 0.6886 | 0.6168 | 0.0718 | 1.1164 |
| DPM | 336.56 | 392.7333 | -56.1733 | 0.857 |
| CS/Minute1 | 5.29 | 5.242 | 0.048 | 1.0092 |
| CS @ 101 | 3.8 | 4.4827 | -0.6827 | 0.8477 |
| Vision Score/Minute | 1.01 | 1.2567 | -0.2467 | 0.8037 |
| Enemy Jungle Camp CS2 | 7 | 9.4993 | -2.4993 | 0.7369 |
| Lane Diff @ 10 | -92.9 | -491.558 | 398.658 | 0.189 |
|
1
Lane minions only, jungle camps are not included.
2
Number of enemy camps taken.
|
||||
table 4
Hunter is definitely an interesting case, since he has roughly the same amount of ratios over 1 and ratios under 1. DPM being lower isn’t too surprising given that Nunu is a supportive tank jungler when several of the other junglers he has played are fairly damage focused like Kindred/Graves/etc. What did catch my eye though was Enemy Jungle Camp CS being lower on Nunu. I expected it to be somewhat higher given that Nunu’s consume helps him be a fairly good control jungler, but maybe Hunter’s playstyle is more objective / gank focused? I’m not sure, and that’s just a loose guess.
| Lobozz Averages | ||||
|---|---|---|---|---|
| Metric | Jinx | Non Jinx Champs | Difference | Ratio |
| Games | 15 | 46 | ||
| Win rate | 0.7333 | 0.52 | 0.2133 | 1.4103 |
| KDA | 3.81 | 3.9943 | -0.1843 | 0.9539 |
| KPAR | 0.6281 | 0.4735 | 0.1546 | 1.3266 |
| DPM | 675.62 | 496.53 | 179.09 | 1.3607 |
| CS/Minute1 | 8.1 | 6.9971 | 1.1029 | 1.1576 |
| CS @ 101 | 72.2 | 70.4586 | 1.7414 | 1.0247 |
| Vision Score/Minute | 0.76 | 0.6721 | 0.0879 | 1.1307 |
| Enemy Jungle Camp CS2 | 2.67 | 1.6229 | 1.0471 | 1.6452 |
| Lane Diff @ 10 | -525.4 | -20.7186 | -504.6814 | 25.3589 |
|
1
Lane minions only, jungle camps are not included.
2
Number of enemy camps taken.
|
||||
table 5
Lobozz is fairly similar to his lane buddy Dean, with all of his notable metrics being over 1 (aside from KDA), with many being significantly higher6. There are 2 notable differences in Win rate, and Lane Diff @ 10 however. Like Hunter, Lobozz holds a positive Win rate even off his signature champ, which his lane buddy Dean can’t say.
Once again I want to look at Lane Diff @ 10, as this is the exact opposite of Dean’s own score. Off of Jinx Lobozz has a fairly even laning phase, but on Jinx he has an average deficit of close to 2 kills worth of gold despite his other metrics suggesting he’d be doing better. Of course Jinx has a poor laning phase which may be partially to blame. Combine that weak laning phase with Dean’s preference for squishier enchanter supports, and the abundance of Nautilus + Thresh, maybe they had to cede some control in lane letting their lane opponents garner advantages elsewhere (eg: plates, roams, etc). It’s hard to really say exactly why this is the case with out more data / the player’s perspectives.
As expected most of the values are above 1 in the Ratio column, there is a reason why they keep playing these champs when they can. But as for the original question of “Are these players still able to perform well off their main champion?”, I would say it depends from case to case. Furthermore, in retrospect I probably should have gotten the over all means for the player’s respective positions in each metric and also compared with them. That way I could say if a score is still “comparatively good”, even off their signature champ. All I can say definitively right now
RebelFox for bringing this idea to me in the first place on Twitter. Pookar for managing to even get and compile all this data in the first place. Always Plan Ahea for letting me interview him to get some crucial insight.
Champion By Player tab in the CSV format so I can load them into R-Studio. + They also have new variables added that weren’t in the original dataset. Furthermore, most existing variables were renamed to be “code friendly”.all role champs 7-27.csv is just a fusion of all 5 role’s individual spreadsheets.all role champs 10g min 7-27.csv is a filtered version of all role champs 7-27.csv with only observations that meet the minimum criteria for a minimum of 10 games in total for a champion, and players who have played at least 10 games of said champion.If Lane Diff @ 10 has 2 negative values, then the interpretation is flipped.↩︎
Enemy Jungle Camp CS, while it is significantly different based off the ratio, I deemed it non-important for everyone except Hunter.↩︎
I’m going off gut feeling when I say “significant” here. A more accurate gauge for this significance would be to look at the Z-scores for both “As champion X”, and “Not as champion X” compared to the sample mean for the player overall. Sadly getting those would be too time consuming for my given time budget, but it is doable though if one tinkered with the Data tab rows from Pookar’s sheet.↩︎
Enemy Jungle Camp CS, while it is significantly different based off the ratio, I deemed it non-important for everyone except Hunter.↩︎
I’m going off gut feeling when I say “significant” here. A more accurate gauge for this significance would be to look at the Z-scores for both “As champion X”, and “Not as champion X” compared to the sample mean for the player overall. Sadly getting those would be too time consuming for my given time budget, but it is doable though if one tinkered with the Data tab rows from Pookar’s sheet.↩︎
I’m going off gut feeling when I say “significant” here. A more accurate gauge for this significance would be to look at the Z-scores for both “As champion X”, and “Not as champion X” compared to the sample mean for the player overall. Sadly getting those would be too time consuming for my given time budget, but it is doable though if one tinkered with the Data tab rows from Pookar’s sheet.↩︎