Soon you’ll notice a “fantasy projections” page now at Vikings Territory at the top (UPDATE: It’s up at the top now!), which is my attempt to provide preseason predictions for every relevant fantasy player in the league. As with most projections out there, there’s a strong statistical approach with rigor and history applied, followed by completely subjective modifications that ruin the whole point of rigorous statistics. Even more precariously, the statistical approach is relatively untested and openly experimental.
I wouldn’t necessarily stake my fantasy league on these projections if I were you, but I think it’s a good experiment to see if a game script-centric approach is one worth pursuing. If you don’t want to read the thoughts behind the process, just take a look at the projections for quarterbacks, running backs, wide receivers and tight ends directly. They should change as the offseason progresses. There’s going to be some interesting numbers in there. If anything seems off, it’s probably misperception on my part, but I’ll just blame the model.
Game scripts are a process pioneered (as far as I can tell) by Chase Stuart at Football Perspective, something I’ve cited several times in the past here. For the most part, what I use them for is an alternate form of power rankings because they provide a different interpretation of readily available data that may be more useful and can suppress annoying noise like “garbage time” and magnify the true importance of close games.
A quick summary: game scripts functionally give us a raw number that tells us the average point differential of a given game over every second of play. That means a team that scores a touchdown early to go up 7-0 and maintains that lead throughout the game will have a higher score than a back-and-forth game where both teams score 30 points until the last five minutes of the fourth quarter, where the winning team scores a touchdown and a gimme field goal to win 40-30.
Because of the nature of comeback wins and whatnot, a team can have a positive game script score and lose the game. What I like about game scripts is that it does a far better job (once you adjust for strength of schedule) in predicting future wins than pure point differential or win percentage.
That’s all relatively academic, however, because Chase’s purpose in developing game scripts was to determine the “real” run-pass balance of a team, where you modified the pass-run ratio according to the average game script a coordinator (or team) had over the course of a season. That’s why, for example, Seattle wasn’t as run-heavy as you might think. They, along with San Francisco, simply led early enough in games to start running out the clock.
It bears out, too. When looking at it from a different angle, Mike Clay found that the Seahawks and 49ers ranked 22nd and 23rd in point differential-adjusted run/pass ratio.
My process was simply to use the “pass identity” scores from the game script data for each coordinator or offensively-minded head coach and regress them slightly to the mean in order to get somewhat more accurate identity scores for 2014.
Volume is so important in fantasy—moreso than rate performance—that accurately predicting production share may be more important than anything else.
Relatedly, a lot of projections out there don’t match up. There are a certain number pass and rush attempts from teams each year, and projections for individual skill players from many services don’t match up to or explain the projections for the team as a whole (the total number of attempts should be higher than the total number of targets because of incidental targets and throwaways, for example. Not always the case).
That may not be a bad thing—it could so happen that predicting performance player by player has a better success rate than predicting from the team level down (which is what I used to do). Because no fantasy expert will get everything right, getting a lot of things as close as possible may mean contradictory reports on team passing yards from quarterbacks and their receivers or finding out that you projected a team to run something like 1200 plays instead of a more reasonable 1050 or so. These things happen with different skews and medians.
In this approach, it was important to track coordinators, not just teams. Because there were something like thirteen new coordinators (McAdoo, Reich, Michael, O’Brien, Lazor, Jackson, McVay, Tedford, Turner, Kubiak, Callahan, Lombardi and Shanahan), it made the job a little difficult. This is really important for figuring out a player like Marlon Brown with the Ravens, who is decent but will be restricted by Kubiak’s relative lack of desire to use slot receivers (in fairness, he didn’t have many good ones at Houston).
The next step was to roughly estimate average point differential for teams heading into next season. For that, I largely used their strength-of-schedule adjusted game script from 2013, and added arbitrary modifiers while also regressing to the mean. In this case, it meant artificially boosting Green Bay, New England and the New York Giants while depressing Cincinnati, San Diego and Baltimore.
Using these new scores, I had rough run/pass ratios for each team, and I modified those based on other factors I was aware of: Cleveland’s excellent stable of running backs but poor roster of receivers, Dallas’ odd coordinator situation, Atlanta’s injured receivers coming back and so on.
After that, determining a number of offensive plays from the pace stats provided by Football Outsiders gave us something to work off of to create a solid number for pass attempts in a game after knowing the ratio.
In order to get more complete fantasy numbers, I then applied net yards per attempt projections onto the expected pass attempts and sacks based off of a three-year sample for most QBs. For the rookies, I used my grades on them (stripping away any “upside” grades) and input rookie years they graded similarly to.
It’s all academic, but from that I’ve created quarterback projections with some assumptions (Cassel starts six games and Teddy nine, while Hoyer/Manziel and Henne/Bortles split the 16 games evenly).
I’ve listed the ADP as per the consensus listed by Fantasy Pros of each player, so there should theoretically be a way to create value when drafting—players projected to get a certain amount with a lower ADP or ranking are likely “steals” to target, while those with a higher ADP than their projection peers are value-suckers to stay away from.
There are a few things about these rankings that make them incomplete:
- No uncertainty bands. The projected point total and yards represent my opinion on their median likelihood in each particular statistic. There are some players with flat curves—“high upside”—like C.J. Spiller, Cordarrelle Patterson, Johnny Manziel and so on. There are also players with a small degree of uncertainty (barring injury) like Adrian Peterson, Calvin Johnson and Peyton Manning. If I continue to write about fantasy, I will identify some of these “flat curve” players.
- I refuse to account for injury, but you might want to. I don’t think injury-proneness is as strong of a trend among individual players that I’d be willing to model it, but if you do, go ahead.
- Projected starters were haphazard. I assumed that none of the rookie QBs will start based on the postdraft chatter from teams, but your mileage may vary. I have Bridgewater starting ten games, and both Manziel and Bortles starting eight games. Running backs and receivers were easier to figure out, but no less subject to interpretation.
DIY These Ranks
You’ll want to modify the projections for your ideal drafting strategy. There are any number of successful approaches out there, and with the popularity of Value-Based Drafting at its peak, alternate drafting strategies (“Upside-Down” drafting, “Zero-RB,” “Zero-WR” and so on) require the ability to correctly identify “flat-curve,” low-median players at a particular position so watch out for those. Almost by design, none of these predictions will be correct, and some will be very incorrect.
Therefore don’t confuse median point projection with draft value.
Generally speaking, uncertainty is bad at the top of a position, great at the bottom of a position, and more good than bad at the middle (because, in the event that a player does poorly, you theoretically have a backup to stopgap the loss). A highly uncertain projection probably breaks even before becoming positive around the late third round of most drafts.
I may do a few posts here and there isolating which players “wide” in terms of potential point distribution in order to identify risks and goldmines. Be wary, though. Outsmarting the fantasy market is very hard.
You’ll also want to model for injury correctly; simply reducing a player’s targets because you think he’ll get injured is not enough; other players need to pick up the slack. In the NFL, because quality has a weak effect on the number of plays, that doesn’t mean a 1:1 substitution. If Gronkowski gets injured and loses half his season, you don’t need to necessarily give 68 plays (carries plus passes) to other players; there may simply be fewer plays to run because Gronkowski would have converted a first down that Edelman wouldn’t have.
You may also want to change the quarterback’s yardage total if you project a skill player’s injury to account for the change in options available to the offense.
Some of these projections are quite a bit off of consensus. I’m not sure if that’s a good thing or a bad thing. I don’t know if it’s reasonable, for example, to project that Russell Wilson will pass 450 times after two seasons of 400 attempts or that RGIII will pass 630 times given that he had 393 attempts, followed by 454 (the model projects that Gruden is more pass-friendly when behind than Shanahan was, so that may be why but that is still generous).
There are also a few glitches; because I worked backwards from targets, not yards, there are many teams whose receivers put together slightly more yards than the quarterback has thrown for, but for the most part, 92-95% of all targets, yards and touchdowns are accounted for.
You’ll also find a few stretches because of the need to assign targets and keep it relatively reasonable. If the Giants are going to throw the ball over 630 times, it’s not unreasonable to assume that Manningham could get over 85 targets, like the deservedly-maligned Reuben Randle and 20 fewer than Odell Beckham (and 50 fewer than Victor Cruz). There aren’t a lot of attractive options for splitting those extra targets other than to give them to Manningham.
Next year, regressing around average depth of target and catch rate over expected depth of target will probably produce more accurate receiver results.
Each team has its own quirks. I would be wary of any confidence for Oakland running backs, for example. For whatever reason, the Oakland fan base is very excited about Latavius Murray and both Jones-Drew and McFadden have injury concerns. Again, if injuries are something you like to pay attention to, then neither MJD or DMC would be worth an investment if the front office is remotely excited about Murray.
One final note: don’t get caught up in the completion rates for the quarterbacks. For teams that have split quarterback situations (Minnesota, Tampa Bay, Jacksonville and Cleveland), I assumed equal completion rates for quarterbacks, merely different YPA rates. This shouldn’t matter for most leagues, but if your leagues do an odd point-per-completion format, then adjust accordingly. I’ve included “points per game” for quarterbacks so you can substitute your own judgments for how many starts you think a QB will get.
If anything, consider this an experiment in backwards fantasy projection—driven by score projection and coordinators more than player history. If you want to use this as your guide, go ahead. It’s nearly impossible to win when you’re playing exactly like the other players. Just know that a) I’m probably not going to play very much fantasy football this year and b) I used a more traditional projection system in past years when I was much more involved and did well with it.