Lastly testing sigma_s evolving over time suggests typical development begins to stabilize after draft year + 1. While I haven't tackled any common criticisms, I have also made no progress understanding what's going on under the hood.

An attempt including interactions between league & age in the linear predictor. Notice the contrast between euro & NA leagues. Could be an interesting topic regarding how selection bias across leagues affect estimates. Maybe most of this is alleviated by including deployment info.

I model age curves for F & D using second order random walks. The distinction between a RW1 & RW2 is cute and helpful for smoothing out some fault lines caused by 20 somethings stuck in juniors. Peak is around 26-27 years old.

Here I produce NHLe? by taking the quotient of the exponentiated league coefficients. Some out of sample eval suggests near constant estimates over seasons (likely due to inappropriate choices on my part).

x_i,u makes more sense! I'll have to correct that. I'm too lazy to show everything this early, but I'll update w/ league coefficients on 2nd draft

Oh I'm just using it as a cover to avoid some statistical jargon I forgot. It's a coefficient for a players effect on point production given the league, their draft age and position. The coefficients are assumed to update each a season as a random walk, which smooths their estimates year to year.

Results seem not great via smell test, wouldn't take it over other point based methods. Uncertainty around talent feels off w/ flat prior. League estimates very sensitive to cutoff choices. Likes the Q a bit too much. Surely some self inflicted issues but also some road ahead beyond plug & chuggin.

Top 20 IN-eligible NCAA D1 Skaters by point per game adjusted for opponent's defence. Lot of exceptional players returning.

Top 10 eligible NCAA D1 Defenders by point per game adjusted for opponent's defence. Teammates Winn and Gosling top the charts.

Top 10 eligible NCAA D1 Forwards by point per game adjusted for opponent's defence [very much a WIP result]. O'Brien had a monster year.

My goal is to build a RAPM model using weights instead of binary indicators for players. At this point a sensitivity analysis w/ synthetic shift data would have to convince me to invest more time into cleaning. I'll be shelving this for now, let me know if you make an attempt as I'm very error prone

The lagging player info can create some potentially unrecoverable issues. Sometimes a player will accumulate more TOI in a segment than possible. At the moment I simply let the excess flow into the prior segment. I also end up with roughly 10min of ice time unaccounted for.

The source provides a game clock but it doesn't seem to update in sync w/ the player info. I ended up with 76 unique game times. 32 of them had multiple TOIs for each player. I filtered conflicting snapshots by comparing them to goalie TOIs calculated from goalie change events in the pbp.

Haven't checked manually yet, hopefully the lag exhibits some predictable patterns.

The chart is just early diagnostics at this point, dots are when the player appears in an event, thickness is players toi / elapsed time in that segment. Not too promising so far on my end at least. I'm curious if others have made similar attempts

Source is from: lscluster.hockeytech.com/feed/index.p...
I'm hitting it every 30 seconds during the game. I think this might be the only way to get some info otherwise destroyed once aggregated.

Yeah the nerds are gunna feast on this. Awesome find!

Maybe Reaves for me. Whoever leveraged the PR machine to make themselves completely infallible to enough people. Probably a good locker room guy too