By the Numbers: AGLA Preseason (Part 1 of 2)

April 5th, 2018 by

Hello Grifballers!

Some veterans might remember “By the Numbers” from an old column I used to write on GrifballHub for the AGLA. The goal of the column was simple – to analyze AGLA stats and use them to identify trends, break down player skill, and predict future performance. With the first season of the AGLA run on PlayGrifball kicking off soon, it’s time to look at some of the returning players this season and see what stories can be told by the numbers.

For this preseason edition of the column, we’ll be taking a look at the players that have signed up so far and their numbers from last season, and also breaking down some of the stats and formulas that we’ll be using throughout the season in these articles. Next week, we’ll continue this analysis by looking at bidding history to try and determine what sort of “fair” credit range we could expect to see for some of these players.

As with any stats-based analysis, its important to note before we go in that stats alone can’t always tell the whole story. There are always little things that players do that can’t be quantified – communication, positioning, and awareness, to name a few – and often those little things can make the difference between a great player and an average one. Furthermore, it’s also important to note that we’re working with a small sample size here – at the time of this writing, only stats from last AGLA (FL17) were readily accessible. With those disclaimers out of the way, let’s dig in.

What Makes a Winning Player?

When I first planned to start writing these articles again, I originally intended for it to be similar to how I’ve written it in the past – use a metric I call “Player Rating” to look at the returning players and identify who the top names and up-and-coming players might be this season.  For those who haven’t seen past iterations of this column, Player Rating (which I will be referring to as PR for the remainder of this article) aims to assign a numeric value to each player that represents overall skill on a scale based on a theoretical “best” player and a theoretical “worst” player.  These “best” and “worst” players are created by taking the worst statistical value (normalized per game or per minute) in each category amongst the pool of players being compared to create a player that would have been either the best or the worst among the given pool of players in every single statistical category being considered.  This creates a rating scale  where that theoretical “best” player would be a 100 and the “worst” player a 0.  Each player being rated then has their stats compared to see how far away they were from those “best” or “worst” players, which is ultimately quantified in the form of PR.  Generally, we see that the top players in the league usually end up with PR numbers somewhere between 50-70.

The problem with PR, however, is determining exactly what stats “matter” – that is, of all the stats that are tracked and can be pulled from the Halo API, which ones have a noticeable correlation to on-court performance?  While the stats used for PR have gone though some tweaks since I first started using, most of those categories have stayed the same: goals, K/D, multikills, stops, and wins.  Each of those stats was chosen for a reason – goals are the objective of the game, higher K/D means its more likely that you are getting kills and stopping or starting pushes, multikills are often what causes a swing in momentum that starts or stops a push, stops lead to turnovers which gets you more possessions, and wins (win percentage, more specifically) helps capture all those “x-factors” that can’t otherwise be quantified.  When I was first doing stats research for this article, I found myself thinking about these categories and their reasons for inclusion.  While all those reasons make sense, I had never really dug in to see whether there was an actual correlation between these stats and win percentage, and that led me to the four-hundred credit question:

What makes a winning player?

To try and answer this question, I took all the stats from last AGLA season and normalized them per-minute, then looked at two subgroups of players.  I first looked at those per-minute stats of players with a win percentage of at least 70%, then looked at those same stats from players with a win percentage of 30% or less and compared the two to see which stats had a noticeable difference between winning and losing players.  The results were somewhat surprising.

Group Min/GM GPG KPM DPM KDR Stops/min MKPM Sprees/min GOPM GDPM
70% or more 15.39 1.12 2.80 2.68 1.01 0.10 0.67 0.05 0.16 0.07
30% or less 16.65 0.46 3.08 3.44 0.85 0.11 0.66 0.05 0.11 0.11

For the most part, the average stats per minute from these two groups were very similar.  Many stats, particularly those dealing with tanking-related medals, were almost exactly the same between the two groups – spree medals per minute for both groups was the same at 0.05 and multikills per minute was almost exactly the same as well, with a difference of 0.01 (1.5%).  To my surprise, the number of carrier kills per minute was also incredibly close, that category also having a difference of 0.01 (9.1%), as were the kill per minute numbers which were separated by 0.28 (9.1%).

Five statistical categories were significantly different between the two groups, all of which had a difference of at least 10%.  Winning players had significantly higher goals per game, with players in that group scoring on average 0.66 more goals per game (143.4% increase).  This higher GPG difference seemed to be consistent across all positions – 60.9% of players in the high win percentage group had a higher GPG than the losing group’s average of 0.46,  and 78.3% of those players had at least 0.4 GPG (only half of players in the low win percentage group hit that same mark).  Only a single player in the winning group (4.3%) had no goals, compared to 26.5% of losing players.  The fact that even non-runners in this group consistently get more goals per game than players in the losing group shows a common trait among winning teams and players: they create more opportunities for their team to score.

Two tanking metrics particularly stand out between the two groups.  Winning players had a noticeably better KDR, netting an extra 0.16 kills per life on average over the losing group, an increase of 18.8%.  This increase in KDR, interestingly, wasn’t due to an increase in kills, however – in fact, the winning group averaged less kills per minute than the losing group.  However, this reduced number of kills also came along with a considerably lower number of deaths per minute.  On average, a player in the winning group died 0.76 less times each minute, a decrease of 22.09%.  To put it in more practical terms, winning players are spending more time on the court, which is significant, as having more players on the court for a team makes it more difficult for the opposition to start a push or to try and flank.  To look at this another way, lets consider how much time a player spends out of the action when they die.  While there’s obviously the respawn time, the player also has to take time to regain their bearings when they spawn and catch up to the action on the court.  While how long it will take for the player to cross the court will vary depending on where the action is happening, let’s assume an average case is the player respawning and running to center court. Let’s make a conservative guess and say that on average, it takes six seconds for a player to get back into the action after being killed.  Using that guess with the DPM numbers, we see that for every minute, players from the losing group are spending 20.64 seconds dead (34.4% of the game) while players in the winning group are only spending 16.8 seconds out of the action (28.0% of the game).  Considering that an average game takes 16 minutes, this means that those winning players are on average getting another minute (61.44 seconds, to be exact) of court time.  In a fast-paced game like Grifball, players can get a lot done in that extra minute.

The final two metrics that saw significant variance between the two groups are goal offenses and goal defenses (both per minute).  In practical terms, these two medals combined act somewhat as a measure of how much time a team or player is spending in the offensive or defensive zone – a player with more goal offenses is likely pushing more than a player with less, and vice versa.  Looking at these two numbers, we see that the winning group on average has 45.5% more goal offenses than the losing group and 36.6% less goal defenses.  Interestingly as well, this number is fairly consistent – only 17.4% of winning players had the same GDPM or greater as the losing group, and only two players exceeded it.  The takeaway is this – winning players tend to spend more time taking push opportunities and getting kills in the opposing teams zone and are less likely to “turtle” in their own defensive zone.

As mentioned previously, the one category that really did surprise me was carrier kills per minute, which was actually slightly higher for losing players than winning players, but there are a couple of reasons that help explain this.  First off, the goal offense/defense difference that was just mentioned indicates that winning players tend to be spending more time in the offensive zone, which generally implies that winning players are spending more time with possession of the ball than players in the losing group.  Since you can’t get stops if you have the ball, this limits the carrier opportunities that these players would have.  I believe the nature of Halo 5 also affects how narrow this difference is.  With passing such an integral part of the game, its much rarer to get turnovers off of carrier kills than compared to Halo 3 or Halo Reach.  A good percentage of turnovers are likely coming off of interceptions, missed passes, or plays where the defender kills the receiver before they can pick up the ball, none of which would be shown under carrier kills.  All three of those situations are very dependent on positioning and awareness, which are both less-quantifiable skills which tend to be more developed in high-skilled players.

A New Rating System

Considering the above findings, we currently have four (technically five) metrics that have shown to be correlated with high win percentage players: goals per game, deaths per minute, KDR, and offensive zone presence (goal offenses vs defenses per minute).  Two of these metrics, goals and KDR, were already accounted for in the original PR system, but the other metrics (stops and multikills) showed to have less correlation.  Just knowing what metrics to look at isn’t enough, though – the other half of the rating system is determining which categories deserve more weight than others.  To determine this, we dug deeper and looked at the per-minute averages of these categories for each of the five percentiles (top 20%, 21-40%, etc), which shows which categories have the strongest correlation with win percentage as you move down the scale.

Two of the four metrics had had extremely high correlation: goals per game and KDR.  KDR was unique in that it was the only category that was always lower as you went down a percentile – among our five percentiles, their average KDR was 1.02, 0.96, 0.93, 0.89, and 0.82.  Goals per game was very close to meeting this criteria as well, the only outlier being the top 41-60% group – that group had a GPG of 0.60, while the next percentile below had 0.68, a narrow difference.  Goal offenses were very close as well, with the top 20% being slightly lower than the top 21-40%, though goal defenses were more of a mixed bag – defenses would consistently go from low to high until you got to the bottom 40%, at which point it started to lower again (likely because less skilled players are less likely to get the kills needed to earn these medals even if constantly trapped in their own zone).  Deaths per minute showed a similar pattern, with numbers low near the top and bottom of the scale and jumping up towards the middle, which similarly is probably due to newer and less skilled players being less involved in games compared to high-level players.

Surprisingly, the two other metrics from the original PR system had a high degree of correlation as well despite our initial findings looking at high-win and low-win percentage players.  For both carrier kills and multikills, there was a drop in the top 20% compared to the next tier below, but outside of that both these metrics would become smaller as you went down in percentile.

Considering these findings, the following metrics and weights are what I’ve opted to go with for the revamped PR system:

-KDR (25%)
-Goals per game (20%)
-Carrier Kills (per minute) (20%)
-Offensive Zone Presence (GO/(GO+GD)) (15%)
-Multikills (per minute) (10%)
-Deaths (per minute) (5%)
-Win Percentage (5%)

Rating the Players

We’re running out of time in this week’s column to do a deep dive into individual numbers, but before we leave off let’s take a look at the system in action using last season’s Pro League:

The Eastbrook 69.51 89% 3.26 2.70 1.04 0.09 0.55 78.33%
ACE B S1lentND1 63.22 59% 0.95 3.55 1.09 0.11 1.05 73.96%
ixGingy 60.63 47% 1.29 3.35 1.14 0.11 0.86 70.18%
OHW Crusader 54.94 56% 1.00 3.68 0.81 0.15 0.53 70.41%
The Piggies 51.42 100% 0.29 3.28 0.99 0.08 0.79 80.23%
Saintz The God 49.72 47% 0.29 3.74 1.23 0.09 1.17 48.44%
DIGITAL PAIN 49.04 47% 0.47 3.53 1.08 0.06 1.01 77.27%
Ezy08 46.23 23% 0.77 4.19 0.96 0.10 0.87 63.33%
Invayda Zim 45.06 23% 0.77 3.58 1.12 0.08 0.94 52.75%
jeakilla85 44.94 23% 1.00 2.92 1.06 0.11 0.72 36.36%
Sabasauros Rex 44.79 47% 0.59 3.47 1.09 0.08 0.83 56.55%
NetsPride15 44.27 89% 0.42 4.15 0.88 0.08 0.86 70.27%
OHW Shad 43.85 53% 1.95 2.43 0.93 0.05 0.47 64.95%
Silva 41.06 47% 1.59 3.26 1.01 0.08 0.77 32.65%
Rage More Nerd 38.90 23% 0.54 3.96 0.90 0.08 0.64 66.13%
NOPLEX 38.87 29% 0.57 3.48 0.79 0.11 0.56 57.89%
OHW Tariq 37.56 56% 0.44 3.60 1.00 0.05 0.80 60.61%
KING RICKXL 37.02 17% 1.25 4.13 0.79 0.09 0.52 55.93%
OHW Cleak 36.72 53% 0.53 3.94 0.87 0.07 0.73 57.76%
Almighty Tycoon 35.37 47% 0.47 3.69 0.92 0.08 0.59 54.40%
InfiniteRainge 35.29 23% 0.54 3.78 1.07 0.07 1.04 32.69%
Ouglaf11 35.08 10% 1.10 2.91 0.90 0.09 0.44 40.00%
the man canon 30.38 40% 0.20 2.79 0.90 0.08 0.34 50.00%
OHW BCN 29.46 0% 0.00 3.12 1.07 0.04 0.75 50.00%
ColeTrain1269 28.92 50% 0.00 3.21 0.98 0.07 0.50 41.67%
IcyViper I 26.05 42% 0.83 4.06 0.76 0.05 0.60 48.53%

Next week, we’ll take a look at the same ratings for last season’s Legacy and Ammy players, and look at using stats to predict player’s credit values. Until then, see you on the court!

« | »