Comments on: Great (and Not-So-Great) Expectations

By: Mountain

Mountain — Sun, 02 Nov 2008 19:28:46 +0000

"tied"

or perhaps some improvement could be obtained by discounting outlier games to some extent

By: Mountain

Mountain — Sun, 02 Nov 2008 06:40:47 +0000

I shouldn't have tried the two together but I'd still wonder about whether a different weight set does better. If not as big a shift as I suggested maybe 5, 3.5, 2.5-3. A small difference but if the goal is nudging the overall result might be worth looking.

By: Neil Paine

Neil Paine — Sun, 02 Nov 2008 04:46:08 +0000

I don't know exactly where Tango's Marcel weights come from (though I remember Bill James coming up with a similar weighting system for projections), but I do know that player performance is a lot more stable from year to year than team performance. While Y-1 might be the only significant variable in predicting teams, think about how much personnel turnover happens from year to year -- between trades, the draft, free agency, etc., a team's "true talent level" can drastically change over the course of 2+ years. Players, on the other hand, largely retain their skillsets from year to year; they may improve certain aspects of their game or get older and decline, but, barring injury, "true talent" doesn't change anywhere near as quickly and dramatically at the individual level as it does for teams. That's why I think the Simple Projection System of weighting performance over multiple seasons are appropriate for individual players, even if Y-1 is the only past season that matters when creating expectations at the team level.

By: Mountain

Mountain — Sat, 01 Nov 2008 17:57:09 +0000

Team regression doesn't necessarily imply a change for players but I wonder whether if these player season weights work better than Marcel's. It could be checked, if it hasn't already been.

By: Mountain

Mountain — Sat, 01 Nov 2008 17:49:16 +0000

If previous year is only year that “really matters in terms of establishing expectations” would that suggest a change from

"Give the 2007-08 season a weight of 5, the 2006-07 season a weight of 4, and the 2005-06 season a weight of 3 and calculate the weighted sum of minutes played."

Give the 2007-08 season a weight of 5, the 2006-07 season a weight of 2-3, and the 2005-06 season a weight of 1-2?

By: Ben

Ben — Sat, 01 Nov 2008 14:40:21 +0000

I hadn't followed all the changes to win shares. I knew decimals were added, but I didn't know it could be negative now. If you just zero out the negative contributors, you probably get good results. Minutes and possessions would, of course, be reasonable too.

By: Mountain

Mountain — Fri, 31 Oct 2008 20:21:16 +0000

If previous year is only year that "really matters in terms of establishing expectations" I wonder what you'd find if you weighted previous year performance by month in a way that gave somewhat greater weight to later months. This would help more closely capture trade impacts, rookie development, coach/team adjustments, etc. I know most recent month or game isn't necessarily more accurate for that reason alone but considering these factors I think it might be worth running some alternatives and see if you can improve the prediction even further. Not immediately sure exactly how much NBA weights conference and division games more to later months but could see that as a mitigating factor but perhaps one that could be adjusted for.

By: Neil Paine

Neil Paine — Fri, 31 Oct 2008 17:09:29 +0000

That's a good idea, adding age as a variable, except I'd prefer to weight by minutes or possessions used (or both), to get at the size of a player's role on his team. I wouldn't use Win Shares, though, because they can be negative for really bad players even if they have big roles. Let's take the 1992-93 Mavericks as an example, since they had several notably negative players by WS playing large roles... If you look at their average age when weighted by minutes played and possessions used, it's basically the same (24.9 by MP, 25.0 by Poss). But if you weight by Win Shares, their average age ends up being 31.2, because they had 7 players with negative WS, including 22-year-old Jim Jackson with the absolute worst single-season WS total of any NBA player for the years we're able to calculate the stat. But I do think minute- or possession-weighted age could improve the model. Between 2 equally bad Y-1 teams, one old and one young, the younger team should obviously be expected to do better in year Y, and right now our simple model doesn't make that distinction.

By: Neil Paine

Neil Paine — Fri, 31 Oct 2008 16:48:26 +0000

Yeah, that's exactly what happens. When I regress SRS_Y-1 and WAA_Y-1 on WAA_Y, the r-squared is still 0.44, because they're so correlated with each other (r-value between SRS and WAA is .96). Actually, in the multivariate regression, the WAA_Y-1 variable isn't even significant at the 5% level, which just reinforces SRS' predictive superiority over straight-up winning %.

By: Ben

Ben — Fri, 31 Oct 2008 15:10:43 +0000

One variable to consider adding would be the previous year's win share weighted average age. That could be one number that might pick up the direction a team's headed in.