The Stats Room: Week 4 QB Projections

First off, a little note on projections. They can vary greatly and owners should have an idea on the variation between projections and the results. For instance, here are the actual quarterback results (95 percent range) grouped by rounded projected results (a bit lower because of injured quarterbacks).

I'm not sure what the most important takeaway is from the data but several pieces are important. First, my projection's range of possible outcomes (at the 95 percent level) is +/- 15 points. A quarterback projected for 15 points is most like to produce somewhere between zero and 30 points. At that range, 100 percent guesswork may be an option for most owners. Hopefully, I can improve these results.

Second, after comparing my values to the standard RotoWire projections, the range of outputs for my values is within the industry norm. With most other projections likely having a similar outcome range, I can see why creators don't brag up their results. Their projections could be the best in the industry, but they're still not acceptably accurate.

Third, and most important, owners might not want to know the inherent risk of each player. They just want to go with their best team and not know their results mostly tied to inherent variation.

With the range of results known, here are the results from last week.

Most of my misses (Brady, Flacco, Brissett, Wilson, Rivers) were also missed by the standard projections. Where my system did better was nailing some guys who the standard system was undervalued like Tyrod Tayler (0.1 points vs. 5.0 points off), Josh McCown (0.2 points vs. 7.2 points off) and Carson Palmer (0.7 points vs. 5.3 points). Who knows how far off the numbers will be next week, but I will take one good week. It may be my last.

Even though my projections work, I think I can keep improving them. While I planned to add defensive values to the equation this week, I made an improvement to quarterbacks instead, in-season target inclusion. While previous season data is useful, the addition of in-season data is a major improvement.

Using data from 1990 to 2016, I found out how much to weight previous and in-season pass attempts to predict the next game's rate. Here's a graph of the results

While preseason information never becomes irrelevant, it is only half of the estimate around Weeks 3 and 4. This three-week span likely applies to in-season data production as well. If a quarterback throws a ton of passes over a couple of weeks, it's important to notice the change and give him a 50/50 mix of old and new data. Don't go all in on historic or recent data. It will be a mix of both. If the trend continues, it can be believed more and more.

By week 10, 80 percent of the pass attempts should be in-season data. The in-season data doesn't completely take over but it's close by season's end.

By adding in the in-season pass attempts, I've created Week 4's quarterback projections.

I went back and looked at the top wide receivers from last week and which projection system did best on these few guys. It wasn't mine.

One piece of information I extracted before moving on was how much to weigh preseason and in-season targets. Here's the information.

The resulting weights are like the quarterback numbers with the in-season data taking precedence around week three. Additionally, the in-season becomes 100 percent of the projection in week 15. In the last couple weeks, all preseason can be thrown out.

I think that I'm able to continue to make progress but for now, I won't be able to advance future with wide receivers until I fix my database issue. Incorporating some opposing defensive values is still a major step for the quarterbacks. I hope to do them for next week. Until them, good luck setting your lineups.