Quite a different graph this time around. The power rankings picks were 11-5 this week – pretty good, although it seems everyone did pretty good this week. The big changes is that Jacksonville and Pittsburgh are both smacked down hard, and are in the bottom half of the rankings again.
There are hardly any beatflukes this week, as the graph instead is simplified by a whole ton of beatloops.
Beatflukes:
NYJ=>NE
SF=>SEA
HOU=>MIA
PIT=>MIA
Interesting beatloops (of the 44 total):
CAR=>BAL=>CIN=>CAR
PIT=>KC=>SD=>PIT
CIN=>KC=>SD=>CIN
BAL=>SD=>DEN=>BAL
TEN=>NYG=>DAL=>TEN
DAL=>IND=>PHI=>DAL
STL=>DEN=>NE=>MIA=>KC=>STL
MIN=>CAR=>STL=>DEN=>NE=>MIN
SEA=>STL=>DEN=>NE=>CHI=>SEA
I like this system, but something is seriously wrong if the Dolphins are placed at the top of the graph. That’s just common sense.
Relative positioning in the graph is mostly arbitrary. The question is the links.
It does seem odd to me that Miami–a team with a sub .500 record–has no beatlosses. In fact, they have two losses regarded as beatflukes!
That’s got to be some pretty impressive wins on their resume…
Here’s the sitch on Miami:
HOU=>MIA is a beatfluke
PIT=>MIA is a beatfluke
Beatloops:
MIA=>DET=>BUF=>MIA
MIA=>CHI=>GB=>MIA
MIA=>CHI=>BUF=>MIA
MIA=>CHI=>NYJ=>MIA
MIA=>KC=>STL=>DEN=>NE=>MIA
So, Miami’s losses to Houston, Pittsburgh, Buffalo, Green Bay, the Jets, and New England are canceled out.
Miami’s wins over Kansas City, Chicago, and Detroit are canceled out.
Chicago’s three beatloops don’t help Miami a ton, and don’t hurt Chicago a ton either, so that seems about even to me. The flukes don’t have a huge impact, either – the net result is that they respectively restore MIA’s direct beatwin to TEN, and DEN’s direct beatwin to PIT.
The KC/NE one seems pretty key. That’s a long beatloop, so that means the relationship there is a little vague – it means that it’s pretty likely that either NE=>MIA or MIA=>KC will reappear, depending on how other games shake out.
I’m increasingly convinced that the right approach is not to throw out all beatloops of the same size, but rather, to throw out the losses that generate the most beatloops.
MIA=>DET=>BUF=>MIA, OK, all those games should be thrown out. MIA=>CHI, on the other hand, is the source of three beatloops and is quite clearly the inconsistent game. It should be gone. Same for the STL=>DEN=>NE chain, which is in a bunch of beatloops.
Best way for me to test that would be to implement it and hook it up to my backtester. I also got your emails, too (thanks!) – I might have time to play around with new ideas like that when the playoffs hit.
I was kind of hinting toward the Miami problem in an earlier posting and someone gave a good response, but it looks like the beatflukes are being implemented in a consistent way.
I mentioned how in the (hypothetical) “How beatloops are resolved” the Min=>NO, NO=>TB=>Min, NO=>Det=>Min ought not to be completely eliminated. Instead Min=>NO ought to be beatfluked away because Min has 2 separate beatpaths (one through Tb, one through Det) back to Min.
Doktarr then clarified what I said by saying “For instance, you could have a situation where 2 teams have 5 common opponents, team A is 0-5 against the common opponents, and team B is 5-0 against the common opponents. If team A beat team B, and there are no other common games, then all 11 games are cancelled out, and team A and team B are considered equal despite 1-5 and 5-1 records, respectively.”
This is obviously a situation of A=>B being a beatfluke. So I ask what if there are 4 common opponents, same answer? I say yes. How about 3 common opponents? The actual Mia=>Chi game above looks like it should be a beatfluke, since Chi has 3 alternate routes to Mia (through GB, Buf, NYJ). I believe Mia=>Chi is a beatfluke and the following should be restored: GB=>Mia, Buf=>Mia, NYJ=>Mia.
Then I say how about 2 common opponents? This would be the case in the “How Beatloops are resolved”. I don’t believe they should all be canceled out. Instead Min=>NO should be counted as a beatfluke because NO has 2 other beatpaths to Min. If you don’t agree with that how can you justify using a beatfluke if there are 3 common opponents. If not that, then what about 4, 5, … ? The line has to be drawn somewhere…
I would also suggest even though it hasn’t happened considering this alternate path thing, that it doesn’t have to be common opponents as long as it has multiple beatpaths. Let’s say hypothetically instead of Mia=>Chi=>GB=>Mia, it were Mia=>Chi=>GB=>=>Det=>Mia. This would still count because Chi has an alternate beatpath to Mia, even though it went through 2 teams to get there.
And lastly, considering argument to give strength to beatwins, I think (through some simple testing) that what will happen is some beatwins will have reduced strength, but it won’t change anything. You will end up with things like this (hypothetical) NE=>NYJ=>Buf with NYJ=>Buf having a power of say .5. So unless people want numbers (strength) next to arrows on the charts, it won’t accomplish anything. A .5 beatwin is still better than no beatwin, and although it doesn’t look as strong as a full beatwin, considering this is a ranking system, not a value system, it won’t change anything in the end. You can’t be half better than someone – either you are or you aren’t. I can show this with a simple hypothetical if anyone doesn’t see it.
At one point I thought it was be interesting to weight beatwins with respect to a team. To take the hypothetical in post #6, if you had NE=>NYJ=>BUF, then NE would get 1.5 beatwins from this path, 1 for the direct win vs NYJ and .5 from the 1 step removed path of NYJ=>BUF. If BUF beat a team that NE didn’t have a shorter path to, NE would get 1/3 beatwin from that segment. And from this same path, NYJ would get 1 beatwin, the direct win vs BUF.
But the more I thought about it, the more I liked the simple system of a win is a win. It would really have to add a lot to make it worthwhile to add something like that.
The one advantage would be if a middle of the road team got a beatwin vs a strong team (that wasn’t somehow looped or fluked away), then the beatwins would be reduced somewhat from not being direct wins. But I don’t think it’s worht it.
I agree with Doktarr
Miami is clearly breaking the entire chart. They’ve brought up a whole bunch of crappy teams into the picture.
the whole “removing beatloops” thing is supposed to get rid of the ‘any given sunday’ type upset, but instead, it seems to be giving them way more importance than is needed.
Again, just because Miami is on the top row, it doesn’t imply that it’s of similar quality to IND, CHI, and KC.
It’s usually not a good idea to make tweaks to the algorithm because of one corner case – the side effects usually don’t justify it. For instance, check out:
ARI=>SF=>OAK=>ARI
ARI=>SF=>MIN=>ARI
ARI=>SF=>SEA=>ARI
Under the suggested Miami logic, we’d remove ARI=>SF, and that doesn’t seem to make as much sense as removing MIA=>CHI
We *do* have functionality built in to remove MIA=>CHI if there’s a beatfluke, but the fact is that CHI hasn’t developed an alternate beatpath to MIA.
The beatloops aren’t supposed to get rid of “any given sunday” upsets. Beatloops are just an indication that the relationship between a group of teams is ambiguous, that’s all. It does not *really* imply that the teams in the beatloop are of similar quality, even though I might have lazily described it that way in the past. It’s more to say the relationship is ambiguous, and to simply punt on the question of the relationship between the teams, relying instead on other information in the graph to make the determination. So there’s no implication that Miami is as good as Chicago. It’s more that Miami’s win over Chicago was a super-quality win AND that there isn’t enough other information in the graph to warrant ignoring it.
It’s still an interesting suggestion, and I might test for it later – there are some questions about when it should be resolved (before beatflukes? after beatflukes?) but it’ll be interesting to see.