Latrinsorm
07-15-2014, 04:39 PM
You hear it all the time. The Thunder's athleticism gives the Spurs problems, the Heat are weak to rebounding, it's all about matchups (http://mustangnews.info/blog/2014/04/23/in-nba-playoffs-its-all-about-matchups/). Meaning: team A has a better record than team B, so we expect team A to beat a generic team more often than team B does, but we expect team B to beat team A because of something not apparent in holistic measures such as win-loss record. Frequently this is also used to dismiss statistical analysis, because matchups are frequently couched in grit, toughness, experience, will to win, etc.
But do matchups actually matter? One way we can find out is to look at head to head record over the past 20 years. When we do we find that in the 161 best-of-seven series where a prediction is made it is accurate 108 times for 67.1% ± 7.4%. Aha! Matchups are right!
Not so fast, pal. We've seen this before: win loss records in any small sample have some correlation with overall records, and the whole point is trying to find out if our smaller sample provides information not already apparent in those overall records. What we should be doing is to find those cases where overall record says one thing and head to head record says the other. When we do that head to head only gets 22 of 47 for 46.8% ± 14.6%... you may as well have flipped a coin. When we compare head to head to non-close record it gets even worse: 22 of 56 for 39.3% ± 13.1%... you'd probably be better off flipping a coin!
Okay, but sometimes a team goes 4-0 against another. Surely that at least is evidence of something? Alas, the only relevant case is the 2007 Nets over Raptors, and a sample size of one tells us literally nothing. 4-0 teams are very successful because they almost always have better records overall, not because 4-0 means anything in itself.
But do matchups actually matter? One way we can find out is to look at head to head record over the past 20 years. When we do we find that in the 161 best-of-seven series where a prediction is made it is accurate 108 times for 67.1% ± 7.4%. Aha! Matchups are right!
Not so fast, pal. We've seen this before: win loss records in any small sample have some correlation with overall records, and the whole point is trying to find out if our smaller sample provides information not already apparent in those overall records. What we should be doing is to find those cases where overall record says one thing and head to head record says the other. When we do that head to head only gets 22 of 47 for 46.8% ± 14.6%... you may as well have flipped a coin. When we compare head to head to non-close record it gets even worse: 22 of 56 for 39.3% ± 13.1%... you'd probably be better off flipping a coin!
Okay, but sometimes a team goes 4-0 against another. Surely that at least is evidence of something? Alas, the only relevant case is the 2007 Nets over Raptors, and a sample size of one tells us literally nothing. 4-0 teams are very successful because they almost always have better records overall, not because 4-0 means anything in itself.