"If the Lions beat the Packers and the Packers beat the Bears then that must mean the Lions can beat the Bears" - As kids this was the constant assessment we were making of our favorite teams (more often than not I was making this about the Leafs in lost playoff heartbreaks).
Background
Analytically what we were doing was not too far from an indirect comparison. But faced with questions of match-ups this becomes more complex. In reality, more often than not these assessments may hold to be true. As we watch more games and match-ups we are better able to assess who we think will win - that's why sometimes predictions in the start of the season are so much more difficult to make. What we are doing is collecting data. Each team in a season is a new unit and all games a network of comparisons.
What if we could analyze that network to try to predict who will win? Could we simulate who is really the current best team? So that is what I set out to do with the NFL.
My recent research (yes I have a day job!) applied a rather novel means of analyzing data called a network meta-analysis (Click here to see my recent work and proof I did actually do this). This type of analysis attempts to use all the data to indirectly compare drugs to each other in how well they work or how safe they are. It applies Bayesian statistics which doesn't assume everything can randomly happen but rather that past information and inform our future predictions. Picture this - we have on the market drugs A and B but both A and B have no comparisons to each other because to gain access to the market they had to only compare to placebo (drug c). This area of work sets out to compare A and B indirectly by using the studies compared to placebo.
This type of analysis can be directly paralleled with a season full of games. As we get more information of results we are able to indirectly compare teams and then can run simulations to try to predict possible outcomes and rankings.
What I did and hope to do (Analysis):
I took all the games of this 2013 season and recorded all the scores. The analysis I completed takes into account not only who wins the games but what the scores are (more importantly the difference in the score). I then ran the simulations and I am able to determine the probability for each team of them being in first to last place and the head to head odds of the win. I plan to compare my analysis (predictions) for the last 3 weeks of the season to random chance (50%) and the Vegas picks. I probably wont do as well as the experts but its worth a shot. Ill start today by posting my power rankings compared to the experts power rankings. Ill post predictions Sunday. The overall goal is to use the season data to predict the playoffs.
NOTE: I did not take into account if the games were played at home and injuries and other major changes. These are obvious limitations to this analysis. I know the purists will argue this and I agree this is just a tool to help make those types of analysis.
Results:
Simulated Power Rankings:
|
|||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| NOTE: Way off is judged by greater than +/-5 spots from ESPN placement. Some interesting notes: 1. This analysis takes into account the season as a whole. So recent improvements and trends are not taken into account. So you see teams in some ways higher or lower based on the most recent games in the expert opinion. 2. The analysis weighs blowout difference heavily. So a team like Houston that had 2 wins but many close loses will perform better than expected. On the other hand a team such as New England that has had many close wins (and an upset to bad teams) will not perform as well as expected. Also getting blown out will hurt your ranking. 3. Some surprises in the rankings are : Arizona. New England, Indianapolis, and St. Louis. These moves are easily explained by point 2. 4. The top and bottom of the packs are easy to predict. The center portions are much more clumped. I will need to work on a better algorithm to trudge through the data for the ranking. Perhaps some sort of weighing. This hurts some teams that get upset by low ranked teams and makes their values across the board flat and hard to predict. A great example is Detroit who was hard to place due to their loses to low teams and wins against higher ranked teams (no bias here). |
|||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
No comments:
Post a Comment