outcomes

with half/half training and testing

error


log reg with all features on including win pct and home/away
as low as
train error = 0.3896, validation error = 0.3928

log reg with just points, win pct, home/away
train error = 0.3862, validation error = 0.3690



success

gradient tree boosting with all features including home/away and win pct
test = .622

gradient tree boosting with just points, home/away, and win pct
test = .641

support vector machine with rgb kernel and all features
test = .533

support vector machine with rgb kernel and only pts, home, win
test = .570



4/4 vs 5/8 training made no noticeable difference

log reg and gradient boosting performed best

notably, with log reg, using more features did not improve the testing error,
whereas it did with gradient boosting



conclusion

could be made more accurate with better features, including momentum based
features, cyclical features, etc.

many pros only use points in their statistical models, our results support
the legitimacy of this method