Dynamically Rating Players

The more optimal decisions are not those that work best for any team, but those that work best for how a specific team uses that particular player. This is why analytics in football in my mind is very, very far behind the curve of where it needs to be. While it may be possible that people behind the scenes are using this very methodology and just not telling anyone about it, I would like to propose an improvement.

First, you must map out how your team looks currently in terms of playcalling and yardage distribution. You could further break that down into formation and expectation based upon individual playcall. Nevertheless, at some point you want to create a way of simulating a team's results over the course of the game. For more on how to simulate individual games, I went into a very brief primer that you can find here.

Once you do that, I think it's more important that you look at the entire season as an entire unit. You can simulate the probability of injury of particular players and given an injury have a different gameplan if you like, and given the injury you can simulate the duration consistent with the NFL average for that position (or really specific to that player's risk profile if you want to be as accurate as possible but that's another story). This will allow more accuracy. Nevertheless, a simulation of an entire season may cause you to change the playcalling and usage of individual players and how you use them.

I talk about how to simulate an entire NFL season here to also include playoff odds. In reality, playoffs have a particular value, a playoff home game has a particular value, a superbowl win has a particular value and each outcome has a particular value. A general manager probably should model consistent with his individual goals rather than just maximizing superbowl odds, or even the teams value, as a general manager probably values maximizes his chances of "not being fired" in exchange for a slightly less optimal value returned to the owner. That may also mean sacrificing one season in order to improve the odds in a future year. It may also mean creating a low probability of a disaster year and preferring 7-9 to 9-7 a high probability of the time at the expense of missing the playoffs to avoid the 4 wins or less that may get him fired while generating enough playoff hopes to keep fans in the audience. It also may include being more aggressive on home games.

Nevertheless, it's about how pieces work together synergistically to provide value to your team.  I gave a broad example in an earlier post about how a player who always got 2 yards exactly would be of outstanding value if the team only used them in situations where they needed 2 yards or less, or where they could run 2 plays to that player to get 4 yards or less, or 3 plays 6 yards or less, or 4 getting 8 yards or less.

This "always 2 yards" player is an extreme, but the extreme illustrates the point most clearly... Value depends on utility, and to really optimize our team we need to be aware of exactly how we plan to use every individual peice and how it compliments the system or other players.

An aggressive safety paired with another aggressive safety may give up too many deep passes as a result of aggressive angles trying to get interceptions. Two cautious safeties may give up too many yards underneath so the offense consistently moves the ball downfield, and converts first downs. But one of each may be better than 2 average safeties or above average safeties that don't work as well together. Both preventing big plays AND getting lots of interceptions is a much better result.

These things can be factored in IF the model is good enough to adjust yards allowed in situations and turnovers, and the resulting field position and series of plays following any possible combination of possible outcomes weighted to their probability of occurring.

But I digress from this very important concept, but very difficult to illustrate without sounding too technical or too specific.

The point, as it turns out, is to rank how well players increases your chances of winning a superbowl this year, next year, the year after, and the year after.

If I wanted to go ahead and look 3 years down the road (or 10), I could basically duplicate the entire set of simulation to use the simulator to maximize the chances of winning the superbowl over that timeframe (or other goals), with a caveat that if a particular condition is met (such as 5 wins or fewer), the following years are voided and the probability goes to zero. This allows a GM to better manage his risks of getting fired by the owner if certain goals aren't met.

Nevertheless, however the simulator is set up, you have to try to model the player's usage.

One such example is a receiver. A receiver has a particular value on each running play that he is on the field related to his blocking skills. There's some probability that that block is needed because of the play flowing towards his direction, and IF so, THEN the expectation of the play may actually change slightly. The chances of getting 5 and 6 and 7 and 8 and 9 yards and so on change slightly depending on who is out there. So you must factor this in. The receiver also runs routes and adds value IF a certain coverage is such that the player is in a position where a particular player will draw coverage IF he is successful in releasing off the line of scrimmage and running the route correctly, which also is associated with a certain probability.

A receiver also adds value relative to the coverage he influences in terms of playcalling in regards to the play. For example, Randy Moss had a large tendency to force the opponents to bring a safety over to his side which could limit and simplify the opponent's coverage. This made plays easier to read for the QB, potentially took away  a defender away from the line of scrimmage for the runningback, open things up underneath for the other receivers by increasing their chances of catching the ball. Whether or not Randy did anything after that was additional value, but even before the ball was hiked just being out there he added tremendous value. You have to come up with a way of calculating how much coverage appears as a result of players WHEN they are on the field.

Then there's players getting open which a certain percentage of the time results in a pass attempt and has a range of outcomes associated with it. You have to simulate every play with that player in adn compensate for how things change compared to an alternative to properly value him.

If you can go in on every single play you will ever project to possibly run during a season and you set up the odds of running them or conditions in which you run them and use a simulator, you can determine EXACTLY what a player is worth to your team given your assumptions.

Of course assumptions are going to be wrong, but you can even model that uncertainty with some sort of probability that you won't use the player or that this player will get injured or whatever other chaotic sort of variability on the impact that is possible.

In the end, you could maximize your superbowl chances within X years after looking at every single possible set of transactions in free agency and the draft that you could possibly make.

Even the draft could be modeled by percentages to handicap the probability that you get players in each particular round.

As long as you can estimate the probabilities or range of probabilities, you can simulate, compare, and rank accordingly.

The result of all this should be a vastly improved situation if your assumptions are accurate, while preparing for the possibility if they are not as best as possible.

Obviously this is a ton of work, and I am not about to do this for 32 different teams. But if you can use thhis kind of logic, use of programming and simulations, use of draft picks, determining value of those draft picks based upon probabilities that players are around and what that does to your superbowl chances vs the alternative of trading draft picks for players, you can make literally the very best decisions given the information available at the time.

That's basically the goal.

Of course, that's what GMs are trying to do anyways, but by slicing the assumptions down to the smallest possible unit, we can use whatever data possible to try to measure the expectations and sort of adjust our biases of what we think based upon what a model of what we think shows. In the end, it's still a decision about what you think but why is it superior?

Imagine I had a very larger number of dots in a 3 dimensional cube and I asked you to guess how many. You could do what you felt was best randomly and probably be way off. But what if I could instead narrow it down to a very small unit of measurement and you could calculate how many units go into the full? Then you could come up with an accurate estimate provided that represented the average per small unit. Say there were 12 dots high, 12 dots wide and 12 dots long per "cubed unit". 12*12*12= 1728 dots. Say this "unit" makes up 1/10th of the base of one side of the cube or 17280 across the length. Say there a 10 of these new larger length units that can fill the width to make the full square 1/10th high. Or 172,800 dots. Say this also was 1/10th the height or about 1.728million dots in the entire cube. Which do you think has a better chance of being accurate, a random "guess" based upon looking at it, or actually coming up with exact units of measurements and determining how many of those units of measurements you will see?

This of course is a metaphor. You have the type of play (length), the distrbution of yards per play  or what percentage of the time each play makes exactly 1 yard through 100 (width), and the amount of plays in a game (height).  You are using small amounts of data and applying them accurately to measure the whole. But football is more than just a 3 dimensional objects because you also want to measure how one drive effects the opponent's drive and how one series on offense and defense effects the game. Fortunately we have computers for this, and simulations so that we don't have to come up with 1/10th of the game, and instead can come up with a single play which could be 1/100th of a game if each offense has 50 plays.

You may not be exact, by sampling a small size, but it's a far better "guess" than any alternative method, and you can work at adjusting by situations (or to better approximate the situations. Continuing with the dot metaphor, if there are huge chunks of clusters of lots of dots and several with very few, you probably want to factor that in. How many clusters, how many dots per clusters, how many dots per spacious area and so on. IN a football game, adjustments to circumstances provide a more ideal set of decisions that have different results. A run play after the defense has put 9 in the box is going to have different results than if they only have 5 and also depends upon what formation you are running. This is also why variable change and calling audibles is important.

You have to be aware of how new information changes the odds and factor this in to accurately model everything.

If you can model an entire game and measure the approximate odds of winning, gameplan over a season to maximize superbowl victory and then maximize that same superbowl winning gameplan with a different set of players, you have a better way to understand how much each player is worth to you.