Hockey is a game with both luck and skill elements. Due to that luck/randomness element, you can never predict 100% of the results in hockey (and that's a good thing, otherwise there would be no point watching the games). So here's my question:
In the long run, what algorithm will give you the most correct predictions? What is the maximum percentage of playoff series for which you could correctly predict the winner?
Comparison of Prediction Methods
These results are calculated from 2006-2013 (with Fenwick stats calculated 2008 onwards).
Traditional stats
Team with home ice: 69/120 (57%)
Better regular season record: 69/120 (57%)
Better regular season record, ignoring shootouts: 72/120 (60%)
Shot differential: 77/120 (64%)
Goal differential, ignoring shootouts: 81/120 (68%)
New stats
Better Fenwick tied: 51/90 (57%)
Better Fenwick Close: 55/90 (61%)
Better Score-Adjusted Fenwick: 58/90 (64%)
Goal differential has been the best predictor of playoff success in the salary cap era. Among the new stats, Score-Adjusted Fenwick so far has the highest predictive value.
Of course, a major caveat is that we only have 6 or 8 playoff tournaments which isn't a huge sample size.
Theoretical Limit
Here's a study from the Nations Network about the relative impact of luck and skill on results in different sports.
http://nhlnumbers.com/2013/8/6/theoretical-predictions-in-machine-learning-for-the-nhl-part-ii
The researcher found that actual NHL standings are indistinguishable from standings in a league where only 24% of games go to the "better team" and the other 76% of games are decided by a coin toss.
An implication of this is that in the long run, you can't correctly predict the winner of more than 62% (24+76/2) of NHL games. Parity is real.
The 62% number was calculated from regular season results. If that number also holds in the playoffs, then the binomial formula says that the "better" team wins 75% of playoff series.
http://nhlnumbers.com/2013/9/6/machine-learning-predictions-of-playoff-series
Here's a later study from the same author.
After an iterative cycle of training and validating, he was able to reach 74% accuracy.