Unlock More Content Like This With A
Footballguys Premium Subscription
"Footballguys is the best premium
fantasy football only site on the planet."
Matthew Berry, NBC Sports EDGE
Welcome to the first ever edition of a weekly article we're calling, "Matchup Zone." The general framework is that I'll use statistics for the players and their matchups to come up with a points projection for their upcoming games. Then I'll briefly discuss a handful of those projections, either because I whole heartedly agree with what the stats say, I vehemently disagree with them, or I think there's too much uncertainty in the matchup to rely on them. Unlike in baseball, stats can only take you so far in football, so it's always a good idea to use your brain and eyes in addition to your calculator.
Four Commandments of WEEKLY projection
Since this is the inaugural piece of the series, allow me to first spend a few paragraphas explaining my principles for projecting weekly fantasy football points. When embarking on a journey, we need a map to tell us where to go and where not to go. There must be a method behind the madness, and these are my methods.
thou shalt focus have no other factors before the player himself
On average, about 20 percent of a player's fantasy points in a given game has to do with factors related to the player, while the remaining 80 percent has to do with factors related to the game. (Don't ask me how I came up with those percentages; just know it has to do with something called the intraclass correlation coefficient.) Here's the thing, though: Statistics do a much, much, much better job of explaining the player part of the equation than the game part. That's because almost all of that 80 percent related to games is random, while almost all of the 20 percent related to players is not random. In other words, if I asked you whether Matthew Stafford or Sam Bradford is going to score more points against the same opponent, you would say Stafford because he averages more points in general, and you would be right the vast majority of the time. However, if I asked you whether or not Stafford (or Bradford) is going to score more points against Seattle than he will against San Francisco, you might say San Francisco or you might say Seattle, but you would be wrong more often than you were in the Stafford/Bradford question. Don't believe me? In 2012, Stafford averaged 21.8 points per game, while Bradford averaged 17.2 points per game. Stafford's best game was better than Bradford's best game; his worst was better than Bradford's worst. However, Stafford scored 35.8 points against the Seahawks and only 15.5 points against the 49ers, while Bradford averaged 9.0 points against the 49ers and 6.5 points against the Seahawks. As you see, Stafford did better against both opponents, but one guy did better against Seattle, while the other did better against San Francisco. Go figure.
I assure you that I didn't pick out this example because it conveniently supports my point. This is just the way NFL statistics work: Knowing the player is more helpful for projecting fantasy points than knowing the matchup. It's not that the matchup doesn't matter; it does. If it didn't, I wouldn't be writing this column. We're just going to have to use stats to reliably separate the players, then use some more stats to help give a best guess about the matchup, and then use our brains and eyes for the rest.
thou shalt not overcomplicate things
This one has a couple of components. First, there's a tendency in football analytics to try to be the nerdiest guy in the room, showing off the myriad sophisticated ways one can use to spit out a number. Football isn't business or high finance, though. Sure, we all want to profit from it -- and subscribing to Football Guys goes a long way towards helping you do that -- but byzantine algorithms work much better on Wall Street than on the frozen tundra of Lambeau Field. Truth be told, in putting together the analysis that will serve as the backbone for this column in the future, I ventured into "nerdiest guy in the room" territory. I won't name the byzantine techniques I tried, but suffice it to say that the simplest one of all was the one that worked best. This did not surprise me.
The other way we tend to overcomplicate fantasy football projection is by looking at way too many factors, many of which are of minimal importance. The goal of any analysis should be to find the most accurate projection possible using the smallest amount of information. It's a balancing act for sure, but the trick is to find pieces of information that already incorporate several others, as well as to focus on information that makes theoretical sense. As an example of the latter, some people pay attention to the day and time of a game. Maybe their quarterback has a lower historical average on Thursday night than he does on Sundays. But what explains him playing worse on Thursday nights? Maybe his coaches are really bad at game-planning on short rest? Maybe the opponents' coaches are better at it? Maybe he hates missing his ukulele lessons? Who knows. Information like game day and time may seem important, but you shouldn't rely on it unless you have a good grasp of why it's important.
As an example of needing to collapse many factors into one, there's the standard practice of looking at things like home/away, opponent fantasy points allowed per game, travel, weather conditions, whether or not the player is coming off a bye, and whether or not a player shows up on the injury report. Several of these do in fact help us project points, but that's six things we have to write down or type into a spreadhseet when we can get a good approximation for the same information by using two others: the spread and the over-under. Shockingly, it turns out Las Vegas already incorporates all of the game-related factors into their supercomputers when they set their lines. To boot, any changes they make to those lines are based on the whims of the betting public, so "consensus" is a seventh piece of information we don't have to worry about writing down (or typing in). If that's the case, why do all the legwork ourselves? I'd rather be playing my ukulele.
And, oh by the way, here's another example of it: You relying on Football Guys, where we grind the most important bits of fantasy football information into small bite-size chunks for your consumption. Basically, the idea here is a combination of Occam's razor, "splits happen," and "life's too short to be spending all your free time in front of a computer."
Thou shalt covet predictions, not explanations
Unless you're Biff Tannen or have discovered your own personal wormhole, the only information we have for an analysis of weekly fantasy points is from games that already happened. Because of that unavoidable fact of the space-time continuum, we're always going to be much better at explaining why player scored X number of points in a previous game than predicting how many he's going to score in a future game. (If you don't believe me, compare the accuracy of recap articles to preview articles.) Thankfully though, there's some good news: We can improve how well information from the past predicts the future by (a) modifying our game data to frame them in terms of what the situation was going into the game, and (b) testing our projections on other sets of game data to make sure the factors that are important at one time are also important at other times.
For example, most people think opponent fantasy points allower per game is a useful predictor, which is fine. So let's take the 2011 Saints, who finished that season allowing 21.0 fantasy points per game to quarterbacks. Now, say you wanted to predict how many points they would allow in Week 10 at Atlanta. It turns out that New Orleans' end-of-season average was closer to the 25.5 points they ended up conceding to Matt Ryan and company than was their 19.6 average through Week 9. So the end-of-season average was a better predictor, right? Wrong! Although it was less accurate, the Saints' average after Week 9 is still the better predictor because it actually predicted something. Their end-of-season average was only closer because Ryan's score was already baked into the cake via simple arithmetic.
What we're really trying to say when we're using points allowed to make a projection is that it causes quarterback fantasy points to some extent. If we use end-of-season averages to test this, we're getting things backwards because, in the case of the 2011 Saints, Ryan's score in Week 10 is what literally caused their 21.0 end-of-season average, not vice versa. The only projections it would have been useful for were Matthew Stafford's points in the Wild Card round, Alex Smith's points in the Divisional round, or New Orleans' points allowed the following season.
thou shalt remain humble in thine uncertainty
This is pretty obvious, but often overlooked. As I said earlier, there's a ton of randomness in football, especially at the game-by-game level. Whether you do or do not use statistics to make projections, it's a mug's game to put on an air of certainty. For that reason, my preference is to tell people what the stats say and what they don't say, as well as to put a number on just how uncertain I am about them. In short, statistics are not gospel; we have to use our brain and eyes.
how to project weekly points for quarterbacks
Over the next few weeks, I'll be presenting the results of my analyses on how to project weekly fantasy points one position at a time; today we'll start with quarterbacks. My reasons for going through them one at a time are twofold. First, each one really deserves its own discussion. Second, if I put all of them into one column, said column would be about 10,000 words longer than your attention span probably allows. (I'm guessing this one's already approaching that limit.) Most importantly, though, player and game statistics for 2013 aren't all that reliable yet. Here's a table showing how much average points after X games deviates from end-of-season average for both quarterbacks and points allowed to quarterbacks:
Games Played | QB Error | DEF Error |
1 | 7.12 | 7.28 |
2 | 4.49 | 5.10 |
3 | 3.77 | 3.91 |
4 | 3.17 | 3.29 |
5 | 2.57 | 2.83 |
6 | 2.10 | 2.47 |
7 | 1.81 | 2.08 |
8 | 1.65 | 1.86 |
9 | 1.51 | 1.72 |
10 | 1.47 | 1.55 |
11 | 1.16 | 1.31 |
12 | 1.02 | 1.08 |
13 | 0.91 | 0.89 |
14 | 0.74 | 0.75 |
15 | 0.56 | 0.53 |
As you can see, what we know after two games (aka right now) about quarterbacks and the defenses they will be playing in the future is almost 10 times more inaccurate than what we will know after 15 games (i.e., Row 2 divided by Row 15). Moreover, that inaccuracy doesn't drop below 2.0 points until after Game 6 or Game 7, so we're still about three or four weeks away from having mildly reliable information to work with.
So here's what I did in my analysis:
- Looked at all regular season games from 2007 to 2011 for quarterbacks who (a) finished the season with 200 or more pass attempts, and (b) ranked among the top 24 in fantasy points per game at any point in the season. The goal here is to help you make weekly lineup decisions, and your decision at quarterback in a given week will almost always be between two of the top 24. If you're contemplating starting QB30 this week, my suggestion is trading for a better one instead.
- Calculated all potential factors to be what they were at the time. For Matt Ryan's game against New Orleans in Week 10 of 2011, I used his 19.4-point average going into the game rather than his 20.8 average at the end of the season, I used the Saints' 19.6-point average going into the game rather than their 21.0 average at the end of the season, and so on.
- Used the following standard scoring system: .05 points per passing yard, 4 per passing touchdown, -1 per interception, .01 per rushing yard, and 6 per rushing touchdown.
- Ran a series of multivariable regressions to determine the most predictive equation possible, and then tested the accuracy of that equation on analogous quarterback data from 2012.
Let's start with what turned out to not be important whatsoever. First, quarterbacks gain nothing from bye weeks, whether we're talking about them coming off a bye or their opponent coming off a bye, or various combinations of both. Second, the belief that quarterbacks from teams on Pacific time perform worse in games at 1 p.m. Eastern time is only a myth. Third, quarterbacks don't score fewer points if they were on the injury report heading into the game. At first glance, this may seem surprising, but it's really not when you think about it. NFL starters in this day and age play nearly 0% of the time when they're doubtful, but nearly 100% of the time when they're probable. In the context of making of making weekly lineup decisions, you're never considering a doubtful quarterback, and almost always considering a probable quarterback; all your decision-making relates to "questionable." The interesting thing is that only 57 of the 2,204 games in my data set involved a quarterback listed as questionable, and there was no meaningful difference between their average game and everyone else's. In fact, they actually scored ever-so-slightly more points (14.5 vs. 13.8).
Where things get really interesting, however, is when we look at home/road and opponent points allowed per game. If you look at their influence in a vacuum, you find that they do make a difference. The problem is that they don't make anywhere near as much of a difference as the point spread and over-under, which as I said earlier gives you same information (and a whole lot more). A regression using only home/road and opponent points allowed per game finds that they are statistically significant predictors, but it only has an R2 of 1.4% percent. This means that, while quarterbacks playing at home against a quarterback-friendly defense score more fantasy points than those that aren't, it's only a tiny bit more. Contrast that with a regression that only includes the spread and over-under, which also finds both to be significant predictors, but has an R2 of 13.3%. Essentially, the spread and over-under are nearly 10 times as predictive as home/road and opponent points allowed per game.
Onto what is important, which you no doubt realize by now includes the spread and over-under. All else equal, for every point above (or below) an over-under of 43, an average quarterback scores 0.33 more fantasy points. That may not sound like a lot, but realize for example that five of the 16 games this weekend have over-unders above 48, which means those quarterbacks can be expected to gain nearly two points just by showing up. I don't know about you, but I've lost by fewer than two points plenty of times.
In terms of the point spread, quarterbacks score fewer fantasy points as underdogs than they do as favorites, but the relationship isn't so simple (but still by no means complicated, mind you). Here's a graph:
On the horizontal axis, we have a modified version of the point spread, called implied win probability, and I've modified it so that 0% means a 50/50 matchup. On the vertical axis, we have the difference in fantasy points based on implied win probability above or below 50%. As you can see, there's a curve to the line. This curve means that quarterbacks score more fantasy points when their team is more likely to win the upcoming game, but that gain gets smaller as their team gets more and more likely to win. As a real world example, below is a graph showing Alex Smith's weekly scoring in 2009 according to San Francisco's win probabilities in those games:
Certainly, quarterbacks differ with respect to how much their team's win probability affects their fantasy scoring (although not to a statistically significant extent), but Smith's 2009 season was a very typical case. He did much better when the 49ers were a small underdog than when they were a huge underdog, but he only did slightly better than that when they were a huge favorite. (Wait, the 2009 49ers were actually a huge favorite in one of their games?)
Only two other factors turned out to be important, but they're player-related, not matchup-related. The quarterback's average going into the game was the biggest predictor of all, which isn't surprising given what I said earlier about how much better we can account for differences between players than differences between games. It's worth noting, however, that I'm including David Dodds' preseason projection as a "game" when I compute the player's average. The main reason is that my weekly projections end up being more accurate when I incorporate his projections than when I don't. This makes sense given that Dodds' projections have been more accurate than any other fantasy football expert over the past four years, to the point that they predict a quarterback's end-of-season average better than the quarterback's own average after two games (i.e., the error is lower than what the earlier table shows in Row 2). But also, lest we forget that the Second Commandment of Projection Fantasy Football says I shouldn't overcomplicate things, so if I think projections might be important, why do all the work of creating my own when Dodds has already done it? And he happens to own this very site!
The second player-related factor that helps make weekly projections more accurate is whether or not the quarterback is a newly minted starter, whether that's because of a promotion (e.g., Josh Freeman in 2009), a long-term injury (e.g., Matt Cassel in 2008), or coming out of nowhere to start in Week 1 (e.g., Russell Wilson in 2012). Essentially, this factor fills the unavoidable blind spots in Dodds' projections. Cassel's projection in 2008 was 0.36 points per game because he was supposed to be holding a clipboard all season, so it would have been pretty useless for projecting points in his first start in Week 2. The "new starter" variable fixes this.
So just to recap, the least-complicated formula for projecting weekly points for quarterbacks is based on the following things we know going into the week:
- His season-to-date scoring average, which counts David Dodds' preseason projection as a "game"
- Whether or not he's a new starter
- His team's likelihood of winning the game, minus 50%
- The over-under of the game, minus 43 points
And now for the part where I follow the Fourth Commandment, and tell you how accurate the system is (or is not). For the 2,204 games from 2007-2011 that I used to create the formula, the predictions were off by an average of 5.9 points, and correctly predicted whether the player was going to score higher or lower than his current average 60.5% of the time. That's great, but the Third Commandment says none of it matters if the formula performs much worse using an entirely different set of games. Well, luckily it was just as accurate in 2012, with an average miss of 6.2 points and a 59.2% hit rate, and it's been slightly more accurate than that so far in 2013 (6.0 points, 67.9%).
week 3 quarterback projections
Below is the payoff for all this work both in terms of me analyzing stats and you reading this far: projections for the top 24 quarterbacks heading into Week 3. The columns labeled "Player" and "Points" are pretty self explanatory, but the columns from left to right mean (a) how much of the points projection is due to the matchup, (b) the lower end of an uncertainty range for Matchup%, (c) the higher end of an uncertainty range for Matchup%, (d) the lower end of an uncertainty range for the projection, (e) the higher end of an uncertainty range for the projection, and (f) whether or not the player is going to score higher or lower than his current average.
Player | Matchup% | Matchup%LO | Matchup%HI | Points | PointsLO | PointsHI | Direction? |
Peyton Manning | 15.2% | 7.8% | 20.7% | 27.1 | 18.8 | 35.4 | LOWER |
Michael Vick | 13.9% | 12.5% | 15.1% | 25.4 | 18.6 | 32.2 | LOWER |
Aaron Rodgers | 9.9% | 8.8% | 10.8% | 25.3 | 18.5 | 32.1 | LOWER |
Drew Brees | 16.5% | 11.6% | 20.3% | 23.0 | 16.0 | 29.9 | HIGHER |
Robert Griffin III | 9.8% | 8.9% | 10.5% | 22.7 | 16.5 | 29.0 | LOWER |
Colin Kaepernick | 12.4% | 6.0% | 17.2% | 22.4 | 15.3 | 29.6 | LOWER |
Matthew Stafford | 9.5% | 8.7% | 10.1% | 21.7 | 15.6 | 27.8 | LOWER |
Eli Manning | 4.0% | 3.6% | 4.2% | 21.3 | 15.3 | 27.3 | LOWER |
Matt Schaub | 4.4% | 4.0% | 4.8% | 21.2 | 15.2 | 27.1 | LOWER |
Alex Smith | 10.7% | 9.5% | 11.6% | 21.0 | 15.0 | 27.0 | HIGHER |
Tony Romo | 11.0% | 9.1% | 12.6% | 21.0 | 14.9 | 27.1 | HIGHER |
Philip Rivers | -2.0% | -2.0% | -1.9% | 20.8 | 14.9 | 26.6 | LOWER |
Matt Ryan | 2.1% | 1.9% | 2.2% | 20.7 | 14.9 | 26.6 | LOWER |
Sam Bradford | 1.4% | 0.5% | 2.1% | 20.5 | 14.6 | 26.5 | LOWER |
Andy Dalton | 6.5% | 5.7% | 7.2% | 19.8 | 14.1 | 25.6 | HIGHER |
Tom Brady | 8.0% | 3.6% | 11.4% | 19.8 | 13.5 | 26.1 | HIGHER |
Cam Newton | 4.4% | 4.0% | 4.7% | 19.2 | 13.6 | 24.8 | HIGHER |
Russell Wilson | 4.7% | -5.5% | 12.5% | 19.0 | 11.9 | 26.1 | HIGHER |
Joe Flacco | 2.1% | 1.9% | 2.3% | 18.8 | 13.3 | 24.4 | LOWER |
Ryan Tannehill | 3.1% | 2.8% | 3.3% | 18.6 | 13.1 | 24.1 | HIGHER |
Jay Cutler | -3.8% | -3.6% | -4.0% | 18.6 | 13.1 | 24.0 | LOWER |
Andrew Luck | -9.4% | -13.3% | -6.4% | 18.4 | 12.2 | 24.6 | LOWER |
Carson Palmer | -1.3% | -4.6% | 1.1% | 17.8 | 12.0 | 23.7 | LOWER |
Christian Ponder | 1.9% | -1.0% | 4.1% | 17.3 | 11.6 | 23.0 | HIGHER |
As an example of how to read the table, take a look at the row associated with Drew Brees. He's projected to score 23.0 points against the Cardinals, which is higher than his current average of 22.2 points, and his plausible range this week is from 16.0 points to 29.9 points. Of the 23.0 points I'm projecting, 16.5% is specifically based on his matchup, and the range for that percentage is from 11.6% to 20.3%. Basically, the system views Brees as having the most favorable matchup of the week relative to his projected points, with the fifth-highest floor and the fourth-highest cieling.
Finally, as promised, here are a few paragraphs on a handful of projections I'd like to highlight:
Peyton manning
Don't be fooled by the last column. The system predicts Manning will score fewer points than his average mainly because his average is currently off the charts at 32.0 points per game. You may have heard of this phenomenon as regression to the mean. And yet, despite this likelihood, there's still a decent chance that he beats given his matchup against Oakland and the uncertainty inherent in his projection, there's still a decent chance he scores more than 32.0. The other factor that isn't helping him in Week 3 is the diminishing returns associated with Denver being a large favorite. While I said that quarterbacks are affected by the point spread in somewhat different ways, Manning's recent career suggests he's pretty typical: He's averaged 20.8 points in seven games since 2007 playing as a 13.5-point favorite or better.
JAY CUTLER AND PHILIP RIVERS
Here are two quarterbacks I'd stay away from this week. Both have been playing well above their preseason projections, and both are playing in games with small point spreads and low over-unders. Therefore, the system thinks they'll score below their current averages this week. Rivers' opponent, Tennessee, seems to have found a viable pass defense in 2013, both in terms of pass rush and in coverage, while Cutler's opponent, Pittsburgh, is suffering from an implosion on offense, not defense.
russell wilson
Wilson might seem like a lock for a huge game at home against Jacksonville, but this system views him as the "Reply Hazy" Player of the Week, awarded to the player with the most uncertain matchup. The 18% difference between "Matchup%" and "Matchup%HI" is by far the biggest among the top 24 quarterbacks, and suggests hosting the Jaguars is anywhere from the second-worst matchup to the sixth-best. The problem here, of course, is that a point spread of -19.5 translates to as large an implied win probability as a team can get, and the over-under is only 40.5, both of which don't bode well. Adding to the uncertainty (though not explicitly included in the projection's uncertainty) is that Wilson only has 18 regular season starts under his belt, and has never played a game as this big of a favorite, so I can't tell you about how he specifically has been affected by the line in the past. Currently sitting at QB21 in the season rankings, he's likely to score above his average -- the system agrees -- but I could just as easily see the Seahawks shutting the offense down at halftime or giving Marshawn Lynch and company 50 carries as I can see them running up the score a la their home game against the historically inept Cardinals last season (Also a source of uncertainty: Pete Carroll).