Oct 12, 2008

In-Game Win Probabilities Beta 0.3

Still more improvements to the in-game win-probability model. There are still some bugs (mostly due to the data feed), but all games now get a near real-time graph, and there are separate pages for ongoing games and final games. It considers score, possession, field position, and time remaining. Down and distance adjustments are still in work. Check it out at wp.advancednflstats.com and feedback is more than welcome. It can be pretty mesmerizing. I'll keep it on-line through MNF.
By Brian Burke

7 comments:

  1. This is starting to look a lot like what GridIronMine is doing. For example, here are their WP logs and graphs of this week's games.

    http://www.gridironmine.com/report/nfl/2008/6/

    Obviously, the end game is the most difficult part to get right. The GIM system appears to have made some significant improvements in this area compared to other recent WP models (e.g. ProTrade, Zeus, Krasker).
    ReplyDelete
  2. Very neat. Why do you say they have made significant improvements over the other systems you mentioned?
    ReplyDelete
  3. Tarvaris JacksonOct 13, 2008 08:03 AM
    i think you need to add the brad childress factor. Its no matter the score, time or field pos. you just divide the vikings chances of winning in half.
    ReplyDelete
  4. Tarvaris JacksonOct 13, 2008 08:03 AM
    i think you need to add the brad childress factor. Its no matter the score, time or field pos. you just divide the vikings chances of winning in half.
    ReplyDelete
  5. Brian,

    This is way cool. I'm hooked!
    ReplyDelete
  6. Full disclosure: I have provided a limited amount of technical advice to GridIronMine on their WP system.

    I've studied a number of published football WP systems, including my own primitive version, which was described in one of the old PFP books a few years ago. It gets very difficult to take it down to a play-by-play level without losing accuracy, especially near the end of a close game, where every play can result in a huge swing and where time outs are critical. The challenge is to avoid discontinuities, such as where an obviously good play results in a decrease in WP, and other anomalies that just plain look wrong with no apparent explanation.

    The GIM system appears to have worked out most of these breakdowns. It isn't perfect, though, and I still wonder whether building a full Markov chain will ultimately work better than a regression-based approach to win probability.

    That said, I find it exciting that football WP models are becoming more widespread. Hopefully these applications will make the concept more accessible to the mainstream football fan.
    ReplyDelete
  7. JimA- Thanks for the great tip. I'll definitely keep an eye on GridironMine. Soon I'll describe my own approach in more detail, and I hope you'll have time to comment.
    ReplyDelete

Subscribe to: Post Comments (Atom)