How to compute overall rating from bars

[tzach]

could someone enlighten me on how overall ratings should be computed from player bars?

Barring silly bugs in my code, if I use the weights from Draft Analyzer, it seems that the weighted average does not correspond to what I see in game.

OverallRating = ( Bar1*Weight1 + Bar2*Weight2 + Bar3*Weight3 +....)/(Weight1+Weight2+Weight3+...)


I wonder if this is what other people here do, or if anyone has updated values of the weights that should be used to compute overall ratings from player bars.

[QuikSand]

Long, long ago, a denizen here named ez apparently back-ended this system, and (I think I'm right here) worked out all the linear coefficients like you set up above. That was a long time ago (like maybe FOF 2001), probably at a previous board, and likely un-findable.

As things sit now, I think the variable you may be omitting is the hidden rating(s) for many (all?) positions. If you saw every raw rating, then in theory a pure regression could tease out the coefficients, if in fact they are linear. But with at least one (avoid interceptions, avoid fumbles) missing... I'd think that gets tricky.

I, for one, have given up. I'm comfortable enough with the game now that when I see, frex, a QB with what seems like a too-low overall rating given his apparent bars, I just assume it's being dragged down by a low (hidden) Avoid Interceptions rating. Past that, I basically don't pay much attention to overall ratings, other than the obvious.

[tzach]

thanks Quik -- this makes a lot of sense. indeed, most of the discrepancies that I see in overall ratings are for QBs and RBs, where the hidden bars that you mentioned could be more significant.

i don't care much about overall ratings either, but i'm updating my code to track that as a function of what happens to specific bars.

[corbes]

Another problem is that the draft ratings and weights alone don't generate a prediction of the plus/minus movement that the player will experience on the way from his "starting future" rating to his "peak current" rating.



It's tedious but not all that hard to generate 50 classes or so of draft ratings and season outcomes in playertracker. From there you can introduce any amount of error you want into the data!

[tzach]

thanks corbes -- very good point about the future ratings vs peak current.

to give some more background and context, i have two tools to estimate future ratings of rookies. the first tool finds past players with similar combines/bars and is perhaps very similar to what other people have used (combine checker?). the second is a bit more predictive, and looks at correlations between combines+bars to future ratings (on the more technical side, i use principal component analysis for that, which is quite interesting).

for both tools, i need to assemble a large draft class to get reliable results, in a similar mold as people have done in the past.

I wonder what is the most efficient way of assembling draft classes together (with bars and combine info), and would welcome advice from the more experienced users. What I'm doing at the moment is the following, which is admittedly cumbersome.

1) FA2: export data/scouting data after the draft class is generated;

2) Rename CSVs to unique name (draft_personal, rookies, players_personal, team_information, players_information), otherwise they would be overwritten in the next season;

3) end of season: export data/scouting data;

4) repeat.

[corbes]

I did it long enough ago that I forget the precise procedure. But I used playertracker for this purpose: export to playertracker after the draft class is revealed to capture the low and high blue bars, export again at some point during the season*, rinse and repeat. At the end of it all, there are some export functions in playertracker that will generate all the information in one unwieldy csv.**

* I'm not sure I recall exactly when this should be done or whether a second export per season is needed?

** I also can't recall whether it was one csv or several that are needed to assemble the final product, but at least the csv manipulating only needs to be done once, as opposed to every season.

[tzach]

thanks once more.

so it seems that it is unavoidable to export at least twice during the season, i.e. FA1-2 and end of season, as I don't care about TC movement. it would have been nice if rookie bars and combines were automatically tracked when computing a 10-year league history, for instance.

[Squirrel]

Hey tzach, you know about this, right?

Comparing Two RB's - IHOF Forums

[tzach]

Originally Posted by Squirrel

Hey tzach, you know about this, right?

Comparing Two RB's - IHOF Forums

i knew i had seen this somewhere -- thanks for pointing me out to this thread, squirrel.

indeed using greg's draft analyzer weights gets me within a couple points from the in-game ratings. Except for positions like QB and RB where hidden bars play a role.

i'm happy with the correlations i'm drawing between bars, combines and performance using Principal Component Analysis -- the overall ratings are useful as a way of quickly seeing how guys with specific bars would look in game. also, it provides a quick way of establishing benchmarks for performance, and helping to decide under which circumstances a 45-rated player could perform similarly to a 55 or 60 rated guy at a given role.