Search the Community
Showing results for tags 'ratings'.
Found 3 results
With the 112,503 dataset or any other datasets or battle outputs ( I think one of the ratings websites uses a datafile that is uploaded and not just an API feed); is there a way to figure out how a person over their population of games rates among total xp out of the 30 people that they match up against. To me a person who on average rates in the top 10% of xp would be a "better" player regardless of win/lose/draw. I think it would also be interesting to see how much this rating would correlate to overall Win chance predictions. A game is determined by the 30 people in the match and over a population of matches the only consistent thing would be the 1 person being measured in relation to the other 29 people. If you on average rank 3rd out of 30 then that seems like it would be a "better" player that would over time impact win chances for the team. It would also be easy to evaluate win % chances for both teams. If one team averages 15 out of 30, and the other rates 7 out of 30, then the team with rating of 7 would appear to be a "better" team. I would love to see some comparisons showing the results of this.
EDIT: Wrong section. Should this be in Core Skills & Mechanics? Different people obvious measure it in different ways (Efficiency Rating, WN*, Win Rate), but what is it really? At least, what is your definition or interpretation? My definition is ability to do that which is required to create a situation which will allow your team to win the game with the capabilities of your tank, compared to those against and with whom you battle. If a tank is naturally worse at winning, then that doesn't necessarily mean the player is worse at winning. Tazilon thinks that this is a stupid way of looking at skill. I don't know what the fuck he's on. Does anyone else know? Any idea?
Background Wargaming has not been as good a compiling statistics for players, this is due to either not wanting to give in to player demands like they have done in the past, not wanting bad (paying) players feel worse, or not wanting statistics to expose their woefully balanced tanks. In the mean time, many players were not satisfied with how their win rate reflected the size of their e-peen, wanting to find a measure of how much more they are contributing in game than mere winrates. Thus came the efficiency rating, and WN# family of statistics, the basic premise is to use the limited amount of data Wargaming provided in order to come up with a rating that can predict "player skill". I however, saw many flaws in these rating systems. And would like to come with some summary statistics for players that are free from biases introduced by tank choice. Since wargaming only provide mostly aggregate statistics, the tank one plays is a key influence on the performance. If only 1 type of tank were available in game for example, it is very easy to compare the aggregate stat of all players directly. I also want to come up with a methodology that is both easy to understand and free from arbitrary parameters. Wargaming API One of the key problem of these formulas is the lack of data provided by wargaming. Even though the in game service record provides many details, the online API does not. Here is a run down of statistics provided: 1. Medals earned. 2. General stats: average xp rating win rate max xp 3. Aggregate stats: total spotted total base reset total xp total battles total damage total cap points total kills total wins battle survived 4. Per tank statistics: Battles Wins (Other stats are inaccurate) 5. Clan information Tank Average So what do I mean by that? For each aggregate stat, it should be possible to find a per tank average value through multivariate linear fitting. For example, if we have thousands of player data, each containing information about how often they drive each tank, and one aggregate stat, we can write the following equation: avg_tank1 * battle_tank1 + avg_tank2 * battle_tank2 + ... + avg_tank# * battle_tank# = total This is a simple linear regression, where the independent variables are the battles in each tank, the dependent variable is one of the aggregate stats listed above. After data fitting, we should be able to find the average value per tank of any of the above aggregate stats. So what can you do with the averages? Percentile means that your score is above the given percent of people. People who have taken standard test like the SAT might be more familiar with it. So how do you calculate a percentile from Tank Averages? Given the average values of each tank, for each player we can predict what their particular total should be given the averages, so for example, it could predict given the tanks you drive, you should have a 45% win rate. Ratio = Actual_stat / Expected_stat If we plotted this ratio for all tankers in the data set, it should be a distribution centered around 1. With "better" tankers having a higher ratio, and "bad" tankers having a lower ratio. With this distribution, we can then calculate the percentile score for a particular player in the many differnt categories. Here are the categories I think are important: Winrate TAP, Damage TAP, Kills TAP Here categories that might be interesting: Spotting TAP, Capping TAP, Cap reset TAP, Survived TAP Here is the category that is hard to interpret: Xp TAP Preliminary conclusions If the data can be successfully fitted, this formulation can give the average winrate, damage, kill, spotting, cap, cap reset, survived, and xp earned for every tank in the game, and every player in the game. This information can be compared to the data from vbaddict and xvm server, keeping in mind the sample set of players from each service is different. Vbaddict provides the most accurate information, but probably has the smallest sample of players, XVM has a larger set, but I don't think their server updates very regularly. TAP in theory can have data on a much larger set of players. Data give player rating in individual categories, without using arbituary weights to combine them. It would also give player an idea of which category they are good at, and which they need to improve. Average value for each tank is also helpful in comparing the relative effectiveness of each. Before, only winrates of each tank is available, now it is possible to compare say the average spotting of all scout tanks. However, keep in mind the data not only depend on the tank, but also the type of player who drive it. So if a tank is popular, it would expect to have lower stats, while an unpopular tank is more likely driven by better players. Players who "seal club", will most likely see a rise in all major TAP categories with the potential exception of Damage TAP. Since damage is one factor that scales highly with tier, a player who plays both high and low tiers will see the higher tier weighed more heavily relative to the game played. The average HP for tier 10 is about 20 times higher than tier 1, which means seal clubbing at tier 1 need many games to equal the damage output of 1 tier 10 game. Also notice that the data set of players used to calculate the average most likely will include a lot of "seal clubbers", but not a lot of "seals", thus the Tank Averaged stats for those tanks should be a lot higher than expected. Players who do a lot of Tank Companies will potentially see a much higher winrate TAP, while a much lower kill TAP, and damage TAP. Platooning should improve TAP in all categories, but I suspect it will improve win rate TAP much more than damage TAP and killTAP. Playing too much OP tanks like the M4 in the time of the HEAT buff will not increase TAP by much, as many other players are playing the same OP tank, so the average stats of the OP tank will be increased to reflect that. Implementation The bulk of the method involve fitting data the size of players in the set, which should be in the thousands if not millions. There are stats programs like R's biglm() that does this efficiently. AFter fitting, the actual/expected ratio for each player's stats are calculated, this ratio can then be sorted to find the percentile cutoffs. It also require a website that has the data to do this, I'm looking at you Neverwish.