EDIT: Though my W totals are based on schedule so measuring them against Pyth is probably skewing it slightly?
No, you're next to best, after Crow, in both tables.
I subtracted everyone's 'error' from the '14py' error.
These numbers represent that your guesses look 2+ Wins better than a hypothetical someone who just submitted last year's PythW totals. Bigger is better.
It's a very sketchy method I used this time. But by several variations on the method, our rankings are quite similar.
Subjectively weighted blend of 12 metrics / predictions (mostly here but some from outside including 3 Vegas lines, based on past performance / my impression at 5-15%) then further subjectively tweaked up or down by 1-4, with a few by a little more. About 2 hours to process. If I did 10-50 times the work to adjust things I might do better, break even or do worse.
I really like the basketball reference predictor. Unfortunately, it does not take into account some important things like percentage of games played at home, or back to backs. Over the season these might even eachother out, but this tool is useless after the season ends. Things are really unequal in November/December making this less accurate.
Also - there really should be a confidence interval in regards to sample size. IE after 13/82 games, confidence is X. After 40 games of 82, confidence is Y.
ampersand5 wrote:I really like the basketball reference predictor. Unfortunately, it does not take into account some important things like percentage of games played at home, or back to backs. Over the season these might even eachother out, but this tool is useless after the season ends. Things are really unequal in November/December making this less accurate.
How do you know this? I can't find any further explanation for the basis of that prediction than those words: "Results are based on 1000 simulations of remainder of season."
ampersand5 wrote:
Also - there really should be a confidence interval in regards to sample size. IE after 13/82 games, confidence is X. After 40 games of 82, confidence is Y.
There is a best and worst scenario for each team listed already.
Crow wrote:Subjectively weighted blend of 12 metrics / predictions (mostly here but some from outside including 3 Vegas lines, based on past performance / my impression at 5-15%) then further subjectively tweaked up or down by 1-4, with a few by a little more. About 2 hours to process. If I did 10-50 times the work to adjust things I might do better, break even or do worse.
Ok. Thanks.
Crow wrote:You are also doing well. Custom metric or...?
Individual player rating based on 2014 data with an age adjustment as well as a rookie prediction based on draft position only. Additionally, a wild guess for the minute distribution.
i have bobbofitos as 51 for okc...i thought crow had the lowest prediction at 50
i didn't read too carefully about what you were tracking our predictions against, but it looks like it might be a better method than just current wins + pyth%. if that's the case, i will start using that in the comparison instead.
Two examples where a simulation of the remainder of the season is used. That is a much better approach than using current win% or pyth, because the latter does not take current and upcoming SOS into account.
I think Crow did what I said I did above and what everyone was trying to do, i.e. find a good method to most accurately predict wins. I blend metrics with discretion because I believe that is a good approach, not just because it is easy. I believed and believe that independently; but a certain NBA analyst / rapm producer once told me he used a blended metric approach to come in second in one of the biggest, most prestigious prediction contests in world with one of the hugest cash prizes, so the approach is not unknown or without much merit.
I think it is completely fine what you did and I agree that the likelihood of getting it "right" improves by blending different methods. For my own rating I use a blend of two different approaches (purely boxscore-based plus RAPM-based) as well. Nothing wrong with that.
Anyway, seeing the high correlation to my prediction, any chance it is included in yours, Crow?
Thanks. I thought it was included, but apparently not. Probably because there wasn't a prediction under your name here last season? And I was in a hurry and thought a dozen was enough. It would be weighted if I revised based on performance to date. And if I could set up a machine learning model, I'd let it grind to the best blend, this season and thie last five. And apply that to future projections. But for this, one iteration of a blend may be enough, by a human not a computer. (It was good enough for first here twice before.)
