-
Notifications
You must be signed in to change notification settings - Fork 4
/
README
34 lines (29 loc) · 1.54 KB
/
README
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
This is my submission to the
[kaggle march mania competition](http://www.kaggle.com/c/march-machine-learning-mania).
It was made with the encouragement of the [Columbia Data Science Society](http://idse.columbia.edu/cdss).
The final model I ended up using was gbm. It aggregates a bunch of information,
much of which ended up being pretty irrelevant. In particular I create some
aggregate statistics about each team over the regular season, like win percent,
and margins for winning/losing (surprisingly winning margin is super predcitive).
I use the PlayerRatings library to calculate a glicko rating for each team, and
then calculate a prediction from that (along with raw difference in rating).
I merge in some tournament information, including seed number (obviously very
predictive). Finally I added some of the ranking systems that Nate Silver
[identifies](http://fivethirtyeight.com/features/nate-silvers-ncaa-basketball-predictions/)
as important, Pomeroy, Moore, and LRMC. I couldn't get easy access to Sagarin's
predictor ratings, so I used Sagarin's regular ratings.
The gbm's relative influence information is:
var rel.inf
seedn.diff 45.272478
glickopred 14.791104
winpct.diff 6.241056
wmargin_avg.diff 5.389482
glicko.diff 5.316646
POM.diff 4.761056
SAG.diff 4.173324
wmargin_avg.1 4.018286
MOR.diff 3.604845
LMC.diff 2.514890
margin_avg.diff 2.133449
wmargin_avg.2 1.783383
seedpred 0.000000