A few of the comments have questioned my model because it was based on ESPN's power rankings. For the first few weeks of the season, I think it was a good idea to base the optimization model on the power rankings because what else do you have to go by? It just doesn't make sense to use last year's team statistics or 2009 preseason statistics. So I went with the power rankings which told me that going into week 1 the defending Super Bowl champs Pittsburgh were the best team and the 0-16 Lions were the worst, which made a lot of sense at the time.
However, there are several areas in which using the power rankings leaves something to be desired. For example, what if two teams (such as Tampa Bay and St. Louis) are equally good or bad? Power rankings force one of them to be ranked better than the other. Also, rankings force teams to be given incremental ranks from 1 to 32, but what if the best team is twice as good as the 2nd best team? Power rankings force the model to believe that the best team is as equally likely to beat the 2nd best team as the 31st team is to beat the 32nd team.
So, 3 weeks into the season, I think it's time to incorporate some 2009 NFL team statistics into the model. So before calculating the week 4 picks, I've made a few enhancements to the model. Each team's score (which is the essential input to the model which gets updated every week) will now be calculated based on three metrics:
1. Yards gained minus yards allowed per game (Yard Differential)
2. Point differential per game
3. ESPN power rankings
I began with yard differential / game which is the yards gained per game minus the yards allowed per game. Denver (160 more yards gained than allowed per game) is the best and Cleveland (195 more yards allowed than gained per game) is the worst team at yard differential so far this season. But yard differential / game isn't everything. I wanted to incorporate turnovers and red zone offensive and defensive proficiency which is where the point per game differential comes into play. The Saints are the best team so far in point differential per game, averaging 21 more points scored than allowed through week 3.
Finally, if I would have counted point differential/game and yard differential per game equally, the Saints would have come out as the best team going into week 4. This is why the power rankings are still incorporated into the model, b/c they provide a nice grounding (The saints are now power ranked 4th btw)
So these three metrics are all indexed meaning each of the 3 metrics is re-scaled so that the best team gets a score of 100 and the worst team a score of 0. A team's final score will be their average score across the 3 metrics. Here is how the final score was calculated for all 32 teams (remember, these scores will be re-calcualted every week):
| Team | Yard Differential | Point Differential | ESPN Power Rank | Yard Index | Points Index | Power Rank Index | Final Score |
| BAL | 147.6 | 16.6 | 1 | 96.6 | 89.1 | 100.0 | 95.3 |
| NO | 125.7 | 21.3 | 4 | 90.4 | 100.0 | 90.3 | 93.6 |
| NYG | 146.4 | 10.7 | 3 | 96.3 | 75.5 | 93.5 | 88.4 |
| IND | 90.7 | 9 | 2 | 80.6 | 71.6 | 96.8 | 83.0 |
| DEN | 159.6 | 15.4 | 13 | 100.0 | 86.4 | 61.3 | 82.6 |
| MIN | 57.6 | 10.3 | 5 | 71.2 | 74.6 | 87.1 | 77.6 |
| NE | 132.7 | 3.3 | 7 | 92.4 | 58.4 | 80.6 | 77.2 |
| PHI | 121.3 | 7.3 | 9 | 89.2 | 67.7 | 74.2 | 77.0 |
| NYJ | 59.0 | 10.3 | 6 | 71.6 | 74.6 | 83.9 | 76.7 |
| SD | 60.0 | 3 | 12 | 71.9 | 57.7 | 64.5 | 64.7 |
| DAL | 47.0 | 8.4 | 17 | 68.2 | 70.2 | 48.4 | 62.3 |
| GB | -22.7 | 6 | 10 | 48.5 | 64.7 | 71.0 | 61.4 |
| CHI | 21.7 | 1 | 14 | 61.1 | 53.1 | 58.1 | 57.4 |
| PIT | 56.7 | -1 | 16 | 71.0 | 48.5 | 51.6 | 57.0 |
| CIN | -29.0 | 1.6 | 11 | 46.8 | 54.5 | 67.7 | 56.3 |
| ATL | -78.3 | 1.3 | 8 | 32.9 | 53.8 | 77.4 | 54.7 |
| SF | -43.7 | 4.6 | 15 | 42.6 | 61.4 | 54.8 | 53.0 |
| SEA | 43.6 | 3 | 23 | 67.3 | 57.7 | 29.0 | 51.3 |
| TEN | 16.4 | -4.4 | 18 | 59.6 | 40.6 | 45.2 | 48.5 |
| ARI | -25.0 | -3.7 | 19 | 47.9 | 42.3 | 41.9 | 44.0 |
| WAS | 15.6 | -3 | 26 | 59.4 | 43.9 | 19.4 | 40.9 |
| JAC | -49.0 | -3 | 22 | 41.1 | 43.9 | 32.3 | 39.1 |
| BUF | -71.7 | -2.7 | 21 | 34.7 | 44.6 | 35.5 | 38.3 |
| MIA | -13.7 | -8.7 | 24 | 51.1 | 30.7 | 25.8 | 35.9 |
| HOU | -103.0 | -7 | 20 | 25.9 | 34.6 | 38.7 | 33.1 |
| DET | -97.7 | -9 | 28 | 27.4 | 30.0 | 12.9 | 23.4 |
| CAR | -69.0 | -16.7 | 25 | 35.5 | 12.2 | 22.6 | 23.4 |
| OAK | -143.0 | -7 | 27 | 14.6 | 34.6 | 16.1 | 21.8 |
| KC | -98.0 | -12.3 | 30 | 27.3 | 22.4 | 6.5 | 18.7 |
| TB | -136.0 | -16.6 | 31 | 16.6 | 12.5 | 3.2 | 10.8 |
| STL | -127.3 | -16.3 | 32 | 19.0 | 13.2 | 0.0 | 10.7 |
| CLE | -194.7 | -22 | 29 | 0.0 | 0.0 | 9.7 | 3.2 |
So these are the week 4 scores which will go into the model. Let me know your comments. I will share the model's week 4 picks sometime tomorrow.
Posts
I'm no math genius; but adding this years stats will definitely improve the model. I understand the use of the powerrankings..but what are those rankings based on?...one writers opinion? Now that you use stats; will you also have to incorporate bye weeks when calculating yardage and point differential?
ReplyDeleteNot sure if it would help the model; but it this article talk about using stats to determine different ratios in handicapping games...http://www.thehuddle.com/classics/04_fs_handff_p2.php
And I'm thinking yardage/point differentials will need to be adjusted for bye weeks. Just some thoughts.
The bye weeks will not present a problem since the point differential and yardage differential are each per game metrics. So, for example, next week the Eagles have a bye week, so their point and yardage differential metrics will not change during their bye week.
ReplyDeleteThe power rankings are actually based on an average of several different writers. I believe in relative terms (#1 is better than #2, etc) it is very strong, but uniformly separating each team is a legitimate problem.
ReplyDeleteOne thing that still might be interesting to add is a strength of schedule adjustment...