To hell with basketball. Give me pigskin any day.
This is going to be mostly a re-iteration of what you can find here, but a quick explanation of terms:
Adj Off - scoring offense, adjusted for opponent, and scaled to a 1.00 average. Most teams will fall in the 0.50-1.50 range. The 2013 leader was Florida State (Adj Off = 1.87). The 2013 loser was FIU (Adj Off = 0.29).
Adj Def - scoring defense, adjusted for opponent, and scaled to a 1.00 average. Most teams will fall in the 0.50-1.50 range. The 2013 leader was Alabama (Adj Def = 0.37). The 2013 loser was Eastern Michigan (Adj Def = 1.68).
NOTE: And here ends the entirety of the input used in my ratings. Everything else is either derived from AdjO/AdjD in some fashion (like Adj Eff and Adj Marg below), or represents information for curiosity's sake (like record or "luck"). But at the heart, everything about how I evaluate a team comes down to these two numbers.
SECOND NOTE: I've instituted a cap (of sorts) on scoring. Many rating systems utilize MOV, and have a cap of 24 pts (or so) to help limit outlier performances. I've resisted this before, mostly due to the fact that I evaluate offense and defense separately. That MOV cap isn't as easy to implement here.
But after tinkering around, I think I've got something I can work with. I still don't use a hard cap, but for individual game performances that exceed the 96th %-ile, a logarithmic scaling is involved (for example, Texas A&M originally had a 3.02 offense rating for their game against Alabama in 2013, reduced to 2.69 after this adjustment). I may play around with this more and extend that %-ile threshold, but that's what these are working on right now.
Adj Eff - scoring ratio, adjusted for opponent, and scaled to a 0.50 average. Biased in favor of strong defenses. The 2013 leader was Florida State (Adj Eff = 0.83). The 2013 loser was FIU (Adj Eff = 0.17).
Adj Marg - scoring margin, adjusted for opponent, and scaled to a 0.00 average. Biased in favor of strong offenses. The 2013 leader was Florida State (Adj Marg = 1.49). The 2013 loser was FIU (Adj Marg = -1.11).
Rating - an attempt to merge Adj Eff and Adj Marg and mitigate their respective biases. Scaled to a 0.00 average. Calculated according to the formula:
Rating = (4xAdjE - 2 + AdjM)/2
The 2013 leader was Florida State (Rating = 1.41). The 2013 loser was FIU (Rating = -1.22).
SOS - a simple average of the rating for each opponent. This isn't the only way to calculate SOS, nor is it necessarily the best way. It's simply what I decided to use, mostly because of the ease of implementation. The 2013 leader was Stanford (SOS = 0.44). The 2013 loser was Central Michigan (SOS = -0.41).
C-SOS - a simple average of the rating for each conference opponent. I only track this for my pre-season projections.
Exp Diff - The expected scoring differential between two teams (TmA and TmB), based on their respective ratings. Calculated according to the formula:
Exp Diff = Sx(Rating_TmA - Rating_TmB) + HmAdj
where S is the average pts/gm for the current season (28.3 pts for 2013), and HmAdj is the home field advantage adjustment (currently using 3.25 pts, based on a quick analysis of the 2013 season):
x - Exp Diff; y - Act Diff
NOTE: I've only recently begun to track home/road status for games. Currently, this doesn't affect the actual rating for each team, but it does affect the expected scores, which I use to evaluate my system's performance. It's made a significant impact there. Going back to the 2000 season, simply adding the home/road status for each game and making appropriate adjustments unanimously improved my correlations and absolute error (between expected scoring differential and the actual results).
However, I'm doing this based on the information provided on Sports-Reference. While I've got a decent workflow going, it's a bit tedious (which is why I'm only through the 2000 season) and imperfect. There are probably roughly ten games each season that I'm going to improperly tag as "neutral site" as opposed to the correct home/road status.
Compared to the overall benefit of trying to track home/road status, I'm going to accept these imperfections (and try to work to correct them in time). Just fair warning that a couple games here or there may be incorrect with regards to home field.
"Luck" - the difference between a team's expected win percentage and their actual win percentage (ActW% - ExpW%). 2013's "luck"-iest team was San Diego State (+16%, or +2.1 wins). 2013's un-"luck"-iest team was Temple (-20%, or -2.4 wins).
Expected win percentage is the simple average of win probabilities for each game on a team's schedule, with probabilities calculated from Exp Diff, illustrated below:
NOTE: The model I came up with is designed to mostly align with 1990-2010 data, which is why you see the earlier seasons deviate more. The model still isn't too far off, but fair warning that it's less valid for past seasons. Also, this will change when I input more home/road location data.
NOTE: I have a cap at 99% (occurring at 28.5 Exp Diff). This is probably too low, but I didn't want to have win probabilities of 100% (or more). Also, the model that I use (quadratic) actually has a max of ~98.7%. For the mathematically astute of you, this means I have a discontinuous jump of 0.3% 28.5 pts. I'm okay with this small incongruity, and you'd be amazed how much the model changes with even a 0.00001 change in coefficient values. So the 99% cap is what it is.
Another note on home/road status: I hope this helps clarify the enormous impact that playing on the road vs at home can have on a team's chances of winning. With teams that are reasonably close in quality, this can be a 20% swing in win probability. That's huge! In terms of my ordinal rankings, a team can be expected to punch ~10-15 spots above (or below) their level based on game location.
For example, I currently have Wisconsin vs LSU as a virtual toss-up to open the 2014 season (neutral - 51% Wisconsin). Leaving aside the "neutrality" of a game in Texas (and 2015 in Green Bay), those odds change considerably if the games were on campus:
At Camp Randall (WIS home - 62% Wisconsin)
At Death Valley (LSU home - 59% LSU)
Heck, even for Ohio State vs Michigan (a match-up that isn't really that close according to my projections), the win probability changes over 10% based on where the game is played.
Variance - the variance of a team's weekly performances, according to Adj Off and Adj Def, using the same formula your statistics professor taught you. Actually, I'm just summing the respective variances for Adj Off and Adj Def, which may not be the exact formula I'm supposed to use, but c'est le vie. This should help explain some of the apparent outliers in my ratings, as a highly rated (but inconsistent) team may lose more games than you would expect (for example, Arizona State was 8th in my 2013 overall ratings, but ranking 124th in Variance).
This really isn't anything more than informative, though. Generally, we'd like our teams to be consistent, but you may just be consistently bad. Conversely, a bad team with a wild card may surprise with an out-of-nowhere upset. The "leader" for 2013 was Eastern Michigan (Variance = 0.12). The "loser" for 2013 was Indiana (Variance = 0.62).
Resume - the "BCS-style" rankings I came up with. The only thing that matters is winning and losing, pt margins are ignored. More credit is given for beating opponents with good records, and more demerit is given for losing to opponents with poor records. I spend very little time on these, but I offer them as another way to evaluate each season, with weight on actually winning games. The 2013 leader was Florida State (Resume = 0.50). The 2013 loser was MiamiOH (Resume = -0.58).
I think that about covers it, but feel free to let me know if you want more explanation on things.
So this post is supposed to at least partially reflect some pre-season projections. I'm going to cut your first question off and say no, I don't factor in any recruiting or personnel losses. I simply use the same scoring data that all of my in-season ratings depend on. That's definitely a flaw, and feel free to discount anything I say here because of it. I won't even argue! (well, I won't argue that much, I guess).
Regardless or recruiting and attrition, however, these projections are based on the simple idea that teams continue to be what they are. Since that previous sentence was a lot more vague than I had envisioned, I'll try to lay it down in simple terms:
Bad teams tend to stay bad, good teams tend to stay good.
Obviously, this is not a catch-all. Auburn has certainly forged it's own path, and Michigan State - wait, hold on -
[/washes out mouth]
has shown change is possible. But in general, this is a pretty safe assumption, even in a sport with as much attrition as college football. The key is picking the right things to track. So let's start with the basics.
Going back to 1990 (so... 1991 actually), here is the Y:(Y-1) comparison for each FBS team:
Okay, that's pretty ugly, but the point to take away is that winning percentage is at least a partial indicator for the following season. But this doesn't take into account schedule strength. Obviously, if you trade Alabama for Georgia State, your winning percentage should be affected.
I won't profess that my Resume rankings are perfect, but they do take opponent strength into account:
Hey, a trend! How does that compare to offense and defense specific tracking?
ADJ OFF & ADJ DEF
There's a limit at 0.00 that necessarily skews things:
By the by, I've tried to scale the axes such that they are mostly aligned, but I'm definitely not perfect. Pay attention to the axis labels, is what I'm saying.
And as far as the overall rating?
Finally, I want to address whether "luck" has any inherent skill. It is my opinion that it does not.
There doesn't appear to be any:
The best indicator for how a team will perform this year is how they performed last year. But obviously, that's not the entire case, as fluke seasons can and do emerge for various reasons (witness Auburn 2012 vs Auburn 2013). After some tinkering and fiddling, I use a weight of 0.60 for the previous season (Y-1) and 0.40 for the average of the four years prior to that (Y-2:5). Here's the results (since 1995):
An unsuccessful attempt to answer the questions, comments, or criticisms that might ail you.
Your rankings are wrong. Please don't do this. This is unhelpful, and (to be a bit pedantic) incorrect. My rankings are exactly correct according to the scores I've entered. They would be wrong if I've mis-entered a score (possible, but unlikely due to the number of checks I have). I suspect that the criticism here does not stem from a belief that I've mis-entered scores, at least.
If you are going to criticize my rankings (something I encourage!), I want you to explain why they fall short of your litmus test. I may agree with your chosen litmus, I may disagree. I just ask that you give me your method of comparison, rather than a blanket statement of "they are wrong."
TEAM is too high/low in your rankings. Your rankings are wrong. I like my system. I will still fully admit that it is a simplistic system. Among 120+ teams in college football, outliers are inevitable. While I would like to have a "perfect" system, I realize this is impossible. I try to focus on being "mostly right" instead of "never wrong." I think I do an okay job of that.
If some team is higher/lower than you think is appropriate, I don't want to say I'm right and you're wrong (I'm wrong a lot!). I'd just encourage you to investigate why those differences are, rather than immediately assuming I'm a failure (could still be true, though).
TEAM_A is ranked [x]. TEAM_B is ranked [x +/- 1]. This is exactly the opposite of what I would expect! I'm going to let you in on a secret: I'm over-simplifying, but this critique is the single most annoying thing to answer. Because of things like "math" and how I get Excel to function, I have every team ranked in a strict ordinal set. That's not how reality works, though. If we were being perfectly honest, we'd rate out teams in ambiguous tiers:
But that's a bit unwieldy and unsatisfying to look at. So I sort each team into an ordered list, knowing that said list is a poor representation of reality and that the differences between teams are a lot more gray and blurry than that. If your righteous fury would be abated somewhat by simply shifting a couple teams around a bit, I'd just ask that you remember this first :)
You don't put enough weight on winning or losing. With regards to my regular rankings, I don't put any weight on winning or losing (other than the fact the scoring more points and giving up less points aids in winning). This goes against everything America stands for, I know, but I do have my reasons.
I'm not really concerned about "rewarding" or "punishing" teams for their wins and losses. I'm just trying to get a rough approximation of how good or bad a team is. So when Michigan barely pulled out consecutive wins against Akron and Connecticut last year, I don't think they deserved credit for pulling out close wins against bad teams. The message I took from those games was "Michigan is not a very good team, win/loss be damned."
Secondly, adding a win/loss factor is superfluous in many cases. When Wisconsin plays teams from the Hoosier State, there doesn't really need to be an extra benefit attached to winning the game, because 30+ pt beatdowns pretty much have that covered. So this would really only come into play for close wins and losses, and there's plenty of smart people out there that assert winning or losing close games pretty much comes down to chance. I've looked into this myself, as well.
This may fail to satisfy you. That's totally fine! There are a great many other metrics out there that weight winning/losing more heavily (heck, I've got my Resume rankings, myself). I just want to try and offer some explanation where possible :)
Treating all FCS as a single generic team is flawed. NDSU is a far cry away from Savannah State. Absolutely agree. This is simply due to the limitations of my ability. I don't really have time to track each non-major opponent individually, and certainly not going back the 80+ years of spreadsheets that I have. So my choice was to either drop those games entirely, or create a generic homogeneous blur meant to stand in for those cupcake games. I chose the latter, because with a season as short as college football has, we need all the data we can get. I think this helps more than it hurts, but I readily admit it is a flaw in the system.
Taking into account yards/penalties/rushes/special teams/etc. would make these better. And here's where I have to confess a dirty little secret: I'm not trying to create a "perfect" system. Leaving aside the fact that a "perfect" system is impossible, it may be that adding a bit of yardage data, sprinkled with some winning factors and a dash of special teams yields an optimal rating, whatever that means.
So what? Is that really that informative? To draw a parallel, the biggest flaw with the traditional "quarterback rating" statistic is that it tries to do everything in an attempt to be a "whole-encompassing" statistic. The result is a score on a 0-158.3 scale that really doesn't tell us anything on its own. Can you honestly say what a quarterback rating of 95.0 really means?
Right now, my ratings measure:
1) The ability to score points
2) Preventing the other team from doing the same
There are a couple adjustments along the way, but that's all they are at the core. Easy to explain, easy to understand. I also think that they are pretty decent at picking out the "best" teams each year. Adding other factors in may improve that performance in some ways, but it would suffer a loss of clarity in what my system is trying to measure.
Any other questions? Feel free to ask :)