As much as I look to famous coaches to learn from them, the ones I really get excited about are ones such as Billy Beane of the Oakland A's who was made even more famous with the book "Moneyball". These are the guys who leverage science and mathematics to get an advantage in sport.
The new coach for the Detroit Lions NFL team, Jim Schwartz (this article) , fits into this category as the guys who study the numbers and its relation to the games they're involved with. This area of sport application fascinates me. I hope to make the field one of my research streams as I become a faculty member at a University.
Does it apply to Ultimate? Well, one challenge, as many have discussed before, is the lack of collected data for our sport. Baseball, the favourite for sabermetric research, is a game with many head to head battles that can be quantized. Football has clear stoppages that also can be quantified. Smooth flowing, lower scoring games such as hockey, football (soccer), and Ultimate are harder to quantize.
Flowing sports tend to place more value on systems and human intuition in analyzing what and how to achieve goals. Is there a way to capture this intuition quantitively? I have some ideas, but the unknown makes this area fascinating. Thoughts? Collaborations?
PJ
Wednesday, February 04, 2009
Statistical Coaches to Emulate - is it for Ultimate?
Subscribe to:
Post Comments (Atom)
3 comments:
I've always struggled with ultimate statistics. You have your obvious stats: Assists, Goals, Turns, D's, Points played. But, like you say, so much of ultimate is unquantifiable.
On D, my man never touches the disc, but I don't get a D, this shows up nowhere. The assist to the assist, I throw the swing that sets up the open breakside goal.
Stats I think could really tell something: Touches on offense, Second assists (hockey), Touches against you on D, Pulling percentages (In, distance, Hangtime). Avg stall count per team/player. Time to score.
Ultimate is comparably a young sport. Baseball, one of the oldest, with the most statistically data, has had time to evolve. I think one day, Frisbee will put a greater emphasis on statistics. But, like soccer, so much of what 'good frisbee' is, will not appear on the stat sheet.
I have put a bit of thought into this, and I think that the best (and somewhat tautological) measure of a player's skill is whether they help their team score or not -- that is whether their team is more likely to score if they are playing. Motivated by this, I have written a program that calculates how much an individual player on a team increases (or decreases) their team's probability of scoring by beign on the field. The program, in essence, looks at how the team performed when that player was on the field, taking into account opponent strengths, which other players are on the field, as well as any inherent advantage the team on offense might have. You can also break the stats down by O/D points or O/D possessions instead of looking at overall contribution.
My approach seems to be similar in concept to the adjusted plus/minus stat used in basketball (taking into account that Ultimate is not typically played in discrete points instead of against the clock), but I have not bothered to look up any detailed references on this. The output of the program is something like "Player A increases our team's probability of scoring by 5% if he playes on a given O point and increases our probability of scoring by 15% if he playes on a given D point."
I have seen several posts discussing similar topics elsewhere in the Ultimate comunity, so (when I get some time), I will try to publish my code and a writeup somewhere.
I've thought about applying ideas from what mathy nerds call "combinatorial group testing" to evaluating individual contribution to group performance but it turns out that there's a lot more work on binary group testing rather than the more complicated score-based individual/group testing mentioned in this post.
A probabalistic approach does seem to make sense, and maybe one of these days I'll ask Daphne Koller (or rather one of her students :) what she thinks of it.
In terms of advancing the state-of-the-art if someone posted real datasets of players-on-the-line and scores, folks could then compare different analyses / dimensions / algorithms for computing such stats.
Another idea would be to look not only at how individuals contribute to groups, but to how certain pairs/dyads/cliques work well together (flow anyone?), and use some graph theoretic probabilistic analysis there as well.
Just some non-random thoughts :)
Post a Comment