I’ve been working on some player projections based off Profootballfocus’s player grades.

I have 12 years of data in my sample.

I then use a regression number I formulated, and weight that with the prior 4 years of data.
I grouped players into three groups, so I can weight each differently.

Group 1 - 1-3 years experience
Group 2 - 4 -7 years experience
Group 3 - 8+ years

I use the highest R2 value to determine correct weight/regression amount to add. RMSE may be better though?

My question comes from the result I had with the players in group 1. My weighting system says: y-1*1 y-2*0 y-3*1        

example (gradeY-1*snapsY-1*1+gradeY-2*snapsY-2*0+gradeY-3*snapsY-3*1+regression_grade*regression_snaps)/(snapsY-1*1+snapsY-2*0+snapsY-3*1+regression_snaps)
I found it odd that Y-2 had no predicted value for future success.

The position in this case is QB, so the sample in this group is small n=59

Should I use these values?
If not what would be the correct procedure to use, given my lack of data?

You absolutely should NOT weight T-2 at 0, if you are weighing T-3 at 1.  That’s an overfit.

What I do is follow a decay pattern.  All you have to do is figure out the best decay rate.

For example, start with 0.999^daysAgo, or if you like 0.7^yearsAgo

All you have to do is figure out a best fit for that 0.7 or 0.999 value