Since 2009 I've been working on an NFL power ranking system that takes
into account the strength of schedule in a granular way. Comparing each teams game production in a variety of statistics against its opponents ability in combating that statistic.
Besides being a professional data manipulator, my motivation to do what is commonly described to be "bizarre" or "sick" comes from the lack of effort by so called experts in the sports talk world. Most analysts just look at one stat to make assertions that one team or another has the best offense or worst defense. Usually it's
just total yardage. I don't even thing total yardage is the most importatnt stat to consider when determining a teams prowess in offensive or defensive capabilities.
So I do my own now.
I start by collecting a variety of game statistics including:
rushing yards
passing yards
offensive scoring
offensive touchdowns
defensive scoring
special teams scoring
turnovers
penalties committed
sacks
I calculate each teams rank and average results in these
statistics. Using the teams score as a ratio of the median score of
that category, I can create a weight that will then be applied to the
opposition's score when they play that team. So a defense that plays a
hot offense will have their data point in that category for that game
adjusted downward as a way to compensate having to play a very good
offensive team. It acts to bring all scores to the median.
The
interesting thing is that as a team improves or worsens over the course
of the year, their weight will change; thus changing the scores of
other teams that they played over the course of the entire season. I allow for the weekly readjustment of data points based on the rationale that a team is either good or bad and that the entirety of the season reveals the true nature of that team. It does, however, have an effect of pushing an entire
division up or down over the course of things. Additionally certain
stats aren't weighted in this way as it was determined (by me) that it
was inappropriate for that stat.
Once the teams scores
are adjusted by the calculated weights, those scores then aggregate into
an offensive and defensive ranking calculation. The individual
statistic score is created in a process I call normalization. In many
ranking systems the statistics are just ranked (in this case 1 to 32)
and a team like Cleveland is ranked 25th. So they get a score of 25
where 1 is the best score. But this methodology doesn't take into
account that the results aren't evenly distributed over the sample.
Some teams may be much better or worse, or the gap between the 10th and
11th teams might be bigger than most. So in order to take into account
that the ranking of 1 to 32 isn't a smooth even distribution, I use a
normalization method a fellow I work with taught me not to long ago (he
uses it in County Appraisal analysis). Essentially the method is to
take the difference of the score from the sample's minimum score as a
ratio over the total range. It measures how far a team is from the
lowest recorded result. The highest score (or best) scores a 1 while
the lowest scores a 0. But it allows for all incidences to score at a
level that is reflective of how good or bad the result is in context of
the whole. This method works best on a sample that has a large range
with a wide variety of possible scores.
The
individual score of each team in each statistic is aggregated and then
weighted based on importance. This is were the subjectivity of this
system steps in. For example total points scored by offense is weighted
more heavily than say turnovers. So not only do I determine what is
weighted more or less, but also by how much. In an attempt to take some
of my bias out of the situation I completed a very unscientific poll of
some football fans to determine what stats they deem more or less
important when considering the strength of an offense or a defense.
Once
the normalized scores are calculated and weighted, they are then compiled
into offensive and defensive scores that give you each teams ranking in
overall offense and defense. These rankings (and their subsequent
normalized scores) then pour into the power ranking matrix which
results in the final overall ranks.
So in review,
1) calculate each team weight by category,
2) apply the weight to the opposition scores,
3) normalize each statistical result,
4) aggregate each statistical result with subjective category weight to obtain offense and defense rankings,
5) aggregate offense and defense rankings along with record and penalties to determine power ranking.
Note on penalties - It
seems that taking the time to determine which penalties are offense or
defense isn't all that valuable, it could be made a case that this stat
isn't statistically relevant. I keep it as a sort of tie breaker stat. But it does break out that about half the time the winning team has more or less penalties than the looser.
