Instead of a post with graphs and another that shows minimal error calculations, I'll just show you everything in one long post. Again, we were looking at QB's that had at least 100 passes in each of 2007, 2008, and 2009 (there were 19 such players). We then used 2007 and 2008 stats to predict 2009 performance, and used 2009 performance to find the way to tweak our algorithm to minimize error.
For completions, we found that previous year's stats were much more important than the player's experience level. A formula that used 45% 2007 stats, 45% 2008 stats, and 10% completions percentage by experience level minimized error.
For yardage, it turns out that experience level is just as important as previous years stats. Furthermore, stats from 1 year ago are much more telling that stats from 2 years ago. A formula that uses 15% 2007 stats, 35% 2008 stats, and 50% experience level minimized error. Looking at the graph of yards per attempt vs. experience, we see that players tend to improve steadily until year 12, where we see a step drop-off. The last few years in the graph look good, but these are Brett Favre years only and should be expected of every quadragenerian that goes under center.
Strangely enough, the exact same formula for yardage also applies to TDs thrown. TDs thrown do improve steadily throughout a player's career, until they hit a precipitous decline around year 12, with a Favre-fueled resurgence in year 17.
Finally, interceptions are best predicted by experience level, not necessarily prior performance. What's more, performance from 2 years ago seems to hold more weight then from 1 year ago. Once we have a more extensive database, I expect this trend will reverse. But for now, a formula that uses 6% 2007 stats, 4% 2008 stats, and 90% experience level gives us minimal error.
When we graph interception percent by experience level, we find that while interceptions do drop off quite a bit the first few years, they then level off. There does seem to be a spike around year 14, but that is due to low sample size more than anything else.
Up next, we will figure out how to predict player performance after a trade and any ripple effect that may have on the rest of the team (Donavon McNabb, anyone?)
