APBRmetrics

I am also tracking some public win projections here:

https://docs.google.com/spreadsheets/d/ ... 1209160265

The preseason-only model (which use just preseason results) is dominating.

Should / will EPM's reputation take a hit?

LeBron, RAPTOR?

Not a strong season for me either.

Who has done best / worst at calling the top (10 or 15) team win totals? And the bottom?

Was preseason only DARKO better at one or the other?

Crow wrote: ↑Thu Mar 24, 2022 4:04 am Should / will EPM's reputation take a hit?

LeBron, RAPTOR?

Not a strong season for me either.

I am very surprised to see EPM so far down that list. Were all of the stats-based predictions using similar approaches to minutes projections? That's where Retrodiction (using actual minutes played but preseason player ratings) can really help level the playing field.

DSMok1 wrote: ↑Thu Mar 24, 2022 1:23 pm I am very surprised to see EPM so far down that list. Were all of the stats-based predictions using similar approaches to minutes projections? That's where Retrodiction (using actual minutes played but preseason player ratings) can really help level the playing field.

DARKO uses preseason minutes projections from Basketball Monster. I believe Taylor also used their minutes projections, but am not sure. LEBRON and several other models used minutes projections from Kevin Pelton. I suspect the Basketball Monster projections may be better, given it's part of their core product (season-long fantasy), while Pelton generously puts together and shares his projections as a community service in his spare time.

DARKO also has a bit of an unfair advantage here, since it's a purely predictive model and doesn't need to worry about balancing predictive elements with descriptive. EPM and LEBRON, etc... meanwhile are trying to put together a model that can meaningfully contribute to the MVP discussion, while also being predictive in nature. DARKO's purely predictive focus likely helps here meanwhile.

On using retrodictions - I think it's better to use a shared source of projected minutes rather than using actual minutes. Actual minutes will tend to be suppressed for the best players in blowouts (LeBron rarely plays 40 minutes in a 40 point Laker win), which can create a systemic bias if grading a model based on them.

Generally, while I'm not gonna pretend I'm not thrilled with how well DARKO is performing here, there is just a ton of noise in any single-year projection. The preseason model is doing incredibly well because it was super down on Portland for instance, which was largely plagued by a Damian Lillard injury all year, until they finally decided to pivot and trade CJ. If Dame got healthy in early November, the Blazers could have ended up as buyers instead, etc...

There may also be some oddities this year due to the larger share of minutes played by replacement level players due to covid. It's not clear to me what sort of impact that should have, but it probably factors in somewhere.

For all of those dying to know, yes, I added my Crafted NBA Projections https://craftednba.com/projections to the kmedved spreadsheet and am currently ranked #6.

https://docs.google.com/spreadsheets/d/ ... sp=sharing

Your blend did pretty well. I check it, use it. I call the site out more than the blend itself, though maybe that will change.

The metric average here is slightly higher. Both have DARKO. The metric average has some metrics you don't. Do they help or is it your weighting that slightly hurt your blend?

The performance of both blends makes a strong case for blends. Something I have been behind for many years. Some folks with more visibility and influence could help the case for blends in writing and conversation, if they cared / cared to. They could highlight your blend if they wanted a reliable access blend to point to, use.

I dunno if you want to tweak the blend for the future or how much testing was involved in its formulation. I would try for even better. More weight on DARKO, less on some others? I'd backwards engineer something that beats pure DARKO and see if it holds up going forward. Even if quietly in the background til confident.



I noticed I was tied for 6th on ATS wins. When I miss, I tend to miss by too much. That hurts on MAE. Being more like the pack, at least on half the cases I am "less sure on", might help my results; but I gun for different a lot, gun for the top. Get there occasionally. I'd like it to be more often but it is subject to so much variation / "luck".

I am now more awake to full Darko's large lead over the 2 blends being almost entirely based on preseason. Darko without preseason is just on par with the 2 blends.

I don't know off hand how much other metrics use preseason. If they cared about losing to DARKO they might either monitor DARKO with preseason performance in future years to see if it holds up or start trying to use preseason themselves, new or more. I might have used preseason in some cases but not all and not systematically or that heavily.

What about Darko's method / weighting of preseason is the author willing to reveal?

Note:

While the current blend at Crafted NBA does indeed include DARKO, the blend from before the season did not. Once I realized that RPM was not coming back, I switched over to DARKO in it's place and simplified the formula.

Ok, thanks for the clarification. Originally I didn't think your blend had it but then I saw it in the glossary entry. If I had thought harder I might have understood the backstory on that shift.

The lack of DARKO then hurt your preseason projections and hurt the comparison with metric average. Even without changes, your blend will probably be a strong contender for first next season, though many outcomes are possible.

Crow wrote: ↑Mon Mar 28, 2022 4:00 pm I am now more awake to full Darko's large lead over the 2 blends being almost entirely based on preseason. Darko without preseason is just on par with the 2 blends.

I don't know off hand how much other metrics use preseason. If they cared about losing to DARKO they might either monitor DARKO with preseason performance in future years to see if it holds up or start trying to use preseason themselves, new or more. I might have used preseason in some cases but not all and not systematically or that heavily.

What about Darko's method / weighting of preseason is the author willing to reveal?

I did a basic explanation here: https://twitter.com/kmedved/status/1447024217397022720

The tldr is I use Bayesian padding to blend a prior estimate of each team's strength with their preseason performance. So if I team was projected to be a +3 team, but then played 6 preseason games at a +15 strength, their projection would be (3*X +15*6)/(X + 6). If X is 10, then their projection becomes 7.5. I solve for X to minimize error historically.

I think the current season's excellent preseason results are somewhat of an aberration. There is value to preseason, but I would not expect a preseason-only model to routinely dominate (as it has this year).

Alright, sounds good. Thanks.

dtkavana wrote: ↑Mon Mar 28, 2022 7:37 pm Note:

While the current blend at Crafted NBA does indeed include DARKO, the blend from before the season did not. Once I realized that RPM was not coming back, I switched over to DARKO in it's place and simplified the formula.

I don’t know if you noticed, but ESPN actually (re)released the 2021-22 season RPM player estimates. Also, have you thought about including EPM (Estimated Plus/Minus) in the player rating averages on CraftedNBA? I haven’t seen any error or correlation analysis of DRIP (Daily-Updated Rating of Individual Performance), but what are your thoughts on it since it is the closest comparison to DPM?

Moreover, I saw that you include LA3-RAPM and PREDATOR as a sort of preseason prior. As both retrodiction and prediction tests show that preceding year RAPTOR more accurately reflects future player performance than the PREDAtOR ratings, have you considered using previous year RAPTOR values instead? Additionally, what would you think about using the CraftedPM values from the last year as an alternative as well?

The order isn't changing much if any, but here we still are:This is the worst error in the #1 spot in 4 weeks; and the worst overall avg since mid-Feb.