4 ms·

Invite HN: The Algorithmic NCAA Bracket Challenge

27 points by jraines 18y ago

jraines 18y ago
Mine is going to be based on the "Wages of Wins"-emphasized stats of FG%, FT Attempts, Turnovers, and Offensive Rebounds. If you want to get a little more hardcore, check out HN user lance's post about using Logistic Regression Markov Chains for your bracketology: http://blog.weatherby.net/2009/03/using-lmrc-to-pick-ncaa-tournament.html http://blog.weatherby.net/2009/03/using-lmrc-to-pick-ncaa-to...
- lanceweatherby 18y ago
  The Logistic Regression Markov Chain developed at Georgia Tech to predict results of NCAA tournament games does a good job at picking both upsets and late round game winners. A bit soft in the middle rounds.
tlrobinson 18y ago
Any suggestions on where we can get relevant data to feed to our algorithm? (preferably in a machine readable format. I'd rather not scrape ESPN.com or something) Also, the password to the jottit page you link to isn't "hackernews" like it says.
- jraines 18y ago
  password should be fixed. I haven't found a great site for scrapable stats. I'm using http://msn.foxsports.com/cbk/stats http://msn.foxsports.com/cbk/stats which is at least a step up from ESPN's stats. If anyone has a better source, I'd be very interested as well. And if anyone wants to create one -- I guarantee it will make money from sports touts, especially if you niche down to one sport and charge less than statfox.com
  - jackowayed 18y ago
    pass still isn't working for me.
    - jraines 18y ago
      OK - it's open for anyone now. If not, I've tried again to make the password hackernews
      - jackowayed 18y ago
        nope. maybe jottit is working off of a stale pass or something?
        jraines 18y ago
        OK - screw jottit then. It's set to 'anyone can view and edit' and i've changed the password to hackernews like 5 times. Can someone recommend a similar site? In the meantime -- don't even worry about posting the algorithm. If your bracket starts to kick ass, just email me with your description or a link to your personal blog post about it, and we'll have a writeup of the top performers. Or -- easiest solution yet -- just post it as a message in the Yahoo! group.
        samueladam 18y ago
        http://couch.it/ http://couch.it/ Built on CouchDB by the way.
- te_platt 18y ago
  I've been following my own algorithm for ranking teams since the start of the NBA season. USA Today has the best computer readable format that I could find.
Raphael 18y ago
I went with alphabetical order. Xavier FTW!
- jraines 18y ago
  Funny -- this got downvoted to 0 and is perfect through 5 games
  - Raphael 18y ago
    Sweet!
anon_1111111 18y ago
www.kenpom.com has all the stats you need sons
- jraines 18y ago
  looks very interesting, i'm definitely going to look into this
rms 18y ago
My algorithm: pick the team with the highest seed. In the case of a tie, alphabetical order by city name.
- aneesh 18y ago
  That's probably the best strategy without putting much effort, but I bet some enterprising HN user will come up with a more complex strategy that has a higher expected value.
  - yters 18y ago
    Just do principal component analysis on past data. That'll give the best predictors (based on linear regression) based on the past.
    - aneesh 18y ago
      The hardest part of that is actually finding all the relevant historical data & getting it into a good form. Anyway, I'm not about to spend hours on picking a bracket, so I'll leave that to someone else.
colins_pride 18y ago
We're doing something similar to this at our office this year. The wrinkles are that it isn't a standard tournament bracket point structure, and nobody is sharing their algorithms. The point structure is that each person picks 8 of the 64 teams. For every game that those 8 teams win, the person gets the team's seed added to their point total. This makes the team selection a very, very interesting exercise: pick a high seeded team and be very confident of getting a few points, or pick a lower seeded team and risk getting no points, but perhaps win the whole thing with that one team. I bought a copy (price: one starbucks coffee) of 24 seasons worth of data from a dependable source (he knows i'll poison his next coffee if he made a mistake). Now the issue I'm grappling with is that he and I both know we have the same data, so a game theory element is added. For those who are interested, it turns out that 3 seeds and 6 seeds have historically done fairly well, and I've cooked up some rationalizations that I like for why this might be the case, but then we'll just see what happens this time around.
- jraines 18y ago
  Cool. If you want to enter this one as well, you don't have to share your algorithm. I just think it will be more interesting if at least the winner does share.
grandalf 18y ago
here's my algorithm: 1) first assume that no upsets will occur 2) then apply the upset rule to all brackets and to the final games. upset rule: - teams from the east beat teams from the north and midwest (2 seed points) - teams from the south beat teams from the west (1 seed point) YMMV
te_platt 18y ago
Thanks for setting this up. Should be fun. You'll get my algorithm once the tournament starts :)
gfunk911 18y ago
Algorithm: If ("lower seed's pomeroy ranking" + 0.0085) > "higher seed's pomeroy ranking", the lower seed wins, otherwise the higher seed wins. http://gist.github.com/80978 http://gist.github.com/80978 Feel free to steal my code UCLA beats Gonzaga and Memphis beats Louisville, Memphis beats UCLA for title.
ganley 18y ago
The cleverest dead-simple algorithm I've seen is to favor the team whose head coach has the highest salary. When I first heard of this, it did pretty well that year.

27 points by jraines 18y ago

Mine is going to be based on the "Wages of Wins"-emphasized stats of FG%, FT Attempts, Turnovers, and Offensive Rebounds. If you want to get a little more hardcore, check out HN user lance's post about using Logistic Regression Markov Chains for your bracketology: http://blog.weatherby.net/2009/03/using-lmrc-to-pick-ncaa-tournament.html http://blog.weatherby.net/2009/03/using-lmrc-to-pick-ncaa-to...

lanceweatherby 18y ago
The Logistic Regression Markov Chain developed at Georgia Tech to predict results of NCAA tournament games does a good job at picking both upsets and late round game winners. A bit soft in the middle rounds.

tlrobinson 18y ago

Any suggestions on where we can get relevant data to feed to our algorithm? (preferably in a machine readable format. I'd rather not scrape ESPN.com or something) Also, the password to the jottit page you link to isn't "hackernews" like it says.

jraines 18y ago
password should be fixed. I haven't found a great site for scrapable stats. I'm using http://msn.foxsports.com/cbk/stats http://msn.foxsports.com/cbk/stats which is at least a step up from ESPN's stats. If anyone has a better source, I'd be very interested as well. And if anyone wants to create one -- I guarantee it will make money from sports touts, especially if you niche down to one sport and charge less than statfox.com
- jackowayed 18y ago
  pass still isn't working for me.
  - jraines 18y ago
    OK - it's open for anyone now. If not, I've tried again to make the password hackernews
    - jackowayed 18y ago
      nope. maybe jottit is working off of a stale pass or something?
      - jraines 18y ago
        OK - screw jottit then. It's set to 'anyone can view and edit' and i've changed the password to hackernews like 5 times. Can someone recommend a similar site? In the meantime -- don't even worry about posting the algorithm. If your bracket starts to kick ass, just email me with your description or a link to your personal blog post about it, and we'll have a writeup of the top performers. Or -- easiest solution yet -- just post it as a message in the Yahoo! group.
        samueladam 18y ago
        http://couch.it/ http://couch.it/ Built on CouchDB by the way.
te_platt 18y ago
I've been following my own algorithm for ranking teams since the start of the NBA season. USA Today has the best computer readable format that I could find.

Raphael 18y ago

I went with alphabetical order. Xavier FTW!

jraines 18y ago
Funny -- this got downvoted to 0 and is perfect through 5 games
- Raphael 18y ago
  Sweet!

anon_1111111 18y ago

www.kenpom.com has all the stats you need sons

jraines 18y ago
looks very interesting, i'm definitely going to look into this

rms 18y ago

My algorithm: pick the team with the highest seed. In the case of a tie, alphabetical order by city name.

aneesh 18y ago
That's probably the best strategy without putting much effort, but I bet some enterprising HN user will come up with a more complex strategy that has a higher expected value.
- yters 18y ago
  Just do principal component analysis on past data. That'll give the best predictors (based on linear regression) based on the past.
  - aneesh 18y ago
    The hardest part of that is actually finding all the relevant historical data & getting it into a good form. Anyway, I'm not about to spend hours on picking a bracket, so I'll leave that to someone else.

colins_pride 18y ago

We're doing something similar to this at our office this year. The wrinkles are that it isn't a standard tournament bracket point structure, and nobody is sharing their algorithms. The point structure is that each person picks 8 of the 64 teams. For every game that those 8 teams win, the person gets the team's seed added to their point total. This makes the team selection a very, very interesting exercise: pick a high seeded team and be very confident of getting a few points, or pick a lower seeded team and risk getting no points, but perhaps win the whole thing with that one team. I bought a copy (price: one starbucks coffee) of 24 seasons worth of data from a dependable source (he knows i'll poison his next coffee if he made a mistake). Now the issue I'm grappling with is that he and I both know we have the same data, so a game theory element is added. For those who are interested, it turns out that 3 seeds and 6 seeds have historically done fairly well, and I've cooked up some rationalizations that I like for why this might be the case, but then we'll just see what happens this time around.

jraines 18y ago
Cool. If you want to enter this one as well, you don't have to share your algorithm. I just think it will be more interesting if at least the winner does share.

grandalf 18y ago

here's my algorithm: 1) first assume that no upsets will occur 2) then apply the upset rule to all brackets and to the final games. upset rule: - teams from the east beat teams from the north and midwest (2 seed points) - teams from the south beat teams from the west (1 seed point) YMMV

te_platt 18y ago

Thanks for setting this up. Should be fun. You'll get my algorithm once the tournament starts :)

gfunk911 18y ago

Algorithm: If ("lower seed's pomeroy ranking" + 0.0085) > "higher seed's pomeroy ranking", the lower seed wins, otherwise the higher seed wins. http://gist.github.com/80978 http://gist.github.com/80978 Feel free to steal my code UCLA beats Gonzaga and Memphis beats Louisville, Memphis beats UCLA for title.

ganley 18y ago

The cleverest dead-simple algorithm I've seen is to favor the team whose head coach has the highest salary. When I first heard of this, it did pretty well that year.