New Statistical Calculation

To make a long story short, I thought that some of the calculations currently used for evaluating teams were lacking in one way or another. So, I set about doing my own. These are works in progress, but they are rapidly nearing completion.

Still to come: Comparisons (L20 vs. L20RPI1 vs. L20RPI2) and mad Pairwise rankings!

• RPI - Yes, I know RPI is available. In fact, you can get the most official one from U.S. College Hockey Online. Before, as I noted, my calculation was considerably off compared to theirs. However, with the aid of a mathematic formula, I have recoded my calculation and the two are now identical (Yippee!). The only reason I wanted to calculate my own RPI was so that I could modify the calculation and use the original and modified version in some what-iffing with the Pairwise rankings.
  • My calculated, end-of-year RPI

• L20 RPI - This is part of my "fix NCAA tournament and selection" crusade which will become official when I have finished all my calculations and "what-if" scenarios. Frankly, I think the L20 statistic is extremely unfair, rewarding schools that play a patsy schedule down the stretch and unjustifiably skewering schools that play a more difficult schedule. That reason is, in a nutshell and modified to cover an entire season, the reason that the NCAA uses the RPI instead of straight won-loss record as the overall strength factor. I don't disagree with measuring a team down the stretch, but if that is to be done and the RPI is superior to winning percentage over an entire season, it must be superior over the last 20 games too.

  It sounds simple to implement a "Last 20 RPI," but it really isn't. As you know (go look it up if you don't), the RPI factors in a team's winning percentage at 35% (low because differentials are much higher here than in any of the other categories), opponents weighted average winning percentage (average of weighted winning percentages meaning weighted, divided, then added, rather than weighted, added, then divided - an important distinction, and one I initially failed to perceive) at 50%, and opponents' opponents weighted average winning percentage at 15%.

  The first factor is, despite the lower weight, still the strongest factor due to comparitively large variations in win pct. from one team to another. The second factor is strong, but variations in opponents win pct. are relatively small even over a small (say, 20 game) sample. The third factor is miniscule, though it may make a difference of a place or two, probably in the third or fourth decimal place.

  So why is it difficult to implement a "Last 20 RPI?" Well, after the winning percentage (easy, take wpct. over the last 20 games) you have to decide what basis to use. Take the weighted average winning percentage over the whole season of the opponents of the last 20 games? Or take that percentage only over each opponent's last 20 games? How about the opponents' opponents' winning percentage factor?

  I came up with three calculation methods, but the third would be a nightmare to implement so I won't do it. That would involve a sliding twenty game scale for opp wpct and opp opp wpct, using the opponents previous twenty games to get those numbers (and maybe really going nuts and sliding the opponents' opponents' previous twenty games too). As I said, a nightmare to implement.

  • Method 1 - Follow the link to see the results. In this method, opponents' winning percentage and opponents' opponents' winning percentage are taken over the whole season, with only the opponents of the last 20 games factored into the L20RPI.

  • Method 2 - Only a 20 game sample is taken for each team, and opponents' winning percentage and opponents' opponents' winning percentage is taken only for that sample for each team.

  Of the two, I prefer Method 1, mostly because the earliest of the L20 games was, for some teams, back into December. I don't really think that the end-of-season trends project that well back then and so using only end-of-season results for opponents' winning percentage and opponents' opponents' winning percentage is not preferable.

• Non-conference RPI - Call me silly, call me a troll, but I thought it would be nice to have some kind of statistical measure of which conference was best to use as a factor in tournament selection. In other words, when two teams from different conferences are duking it out for that last spot, who's going to get it?

  Head to head is nice, but with non-conference schedules fairly abbreviated and especially regionalized these days, record vs. common opponents and head to head can be awfully thin. I think it begs for another measure, so I propose this: Take the same RPI formula (yes, I'm going nuts) and apply it to non-conference games between teams of different conferences. This doesn't do much to the won-loss record with only four teams there, though the CCHA gets ahead of Hockey East (oh no!) because the non-conference opponents they played were tougher than the non-conference opponents Hockey East played.

  I seriously think this is a fairly solid measure of conference to conference strength, since it can account for scheduling disparities.