- With the beginning of the NHL season upon us, I take this moment to fill
you in on the status of the APBA disk for this year. I know we like to
get the disk out before the beginning of the NHL season, but that is
not always possible. This year, my work life and some slowly converging
testing sets caused things to drag out. I've done too much work travel
this summer/fall, as my job takes me to really strange places on fairly
short notice (last week was the "lovely" town of Socorro NM). Testing on
a netbook is too slow and painful, so I don't do it on the
road...usually pretty wiped out after working long days anyway. The slow
convergence of the disk was getting annoying, so I sort of "blew up" the
set with a big across the board adjustment and started iterating
again.....went better after that. I was definitely caught in a local
minimum the first run at things.
Good news is, the disk is close to testing/tweaking completion. As in
past years, we have some issues on this disk that I'm trying to iron
out, and I'm starting to hit a plateau of progress. This plateau usually
signals the end of testing/tweaking, as major disruptive changes would
need to be made to get any farther.
Here are the known issues so far for this year:
1) Assists are low for players with high assists/minute stats. This is a
very typical situation that we fight every year (flaw in the engine). I
typically get it to about the same point every year (slope of the
correlation line between real and test of 0.8), and usually can't push
it farther due to the effect of shots on goal on major players. For
example, I have issues trying to get Ovechkin all his shots. If I let
him get all his assists, he comes in 40% low on shots. If I get him all
his shots (which is almost impossible, best I've done so far is 5%
low)...he comes in 25% low on assists. Nothing else I can do. We
typically have about 20-25 players in this boat every year. This year is
no exception. I try to balance the best I can, but I usually want shots
to win out, especially for key goal scorers.
2) For some reason this year, the fights have been difficult to get into
a good correlation. I think that is because I did the process slightly
differently this year. Last year, I worked the fight ratings first, and
then worked the rest of the penalties around that. This year, I worked
the big penalties (majors, fights, misconducts, match, game misconduct)
simultaneously, which makes all of them (except the fights) more
accurate than in the past. The numbers aren't horrible, but the really
high end fight guys are coming in 10-20% low on fights/minute, even with
a full 999 fight rating. There are a ton more of those (999 fight rate)
on this year's disk than in the past.
3). Had to chop some misconduct penalties from high PIM guys. Some of
the high PIM guys (typically goons) were running 200% over their real
life stats in PIM. This was mainly due to minor penalties that get
called, even if their PEN RATE is 0. It is another flaw in the game
we've lived with....high PIM guys get too many minors, even with a zero
PEN RATE. To compensate so there are no ridiculous PIM totals for any
players, I reduced (or zeroed out in some cases) the MISC rate to get
more realistic PIM totals without giving up fights, majors, or game
misconducts. I made an executive decision that we'd rather have a guy
with 4 or 5 misconducts end up with zero rather than have him get 600
PIM. Here is the penalty priority, in order, with respect to
testing/tweaking for accuracy: PIM (essentially minors), fights, majors,
game misconducts, match, misconducts. Gross misconducts have disappeared
from the NHL rulebook, so they are all zeroed out.
Other than that, things look pretty reasonable. I am very happy with the
SOG totals for both teams and players, and the goalies are behaving
better than past years. I'm looking to finalize the testing in about a
week, then I need 3-4 days (less if I can get an 8 hour block of free
time somewhere) to move the ratings from split disk I test with over to
the combined disk. This usually entails bringing over minute weighted
averages of the number ratings from their split cards.
Thus, I hope to have a disk about around October 16th. Please be patient
these last couple weeks....I think it's going to work out to be a
- Hello Dave,Are you still on target for this week-end?Thanks a lot for all the effort, sincerelyGhislainGo Habs Go!
Date: Sat, 30 Aug 2014 10:49:50 -0700
Subject: [hockeydisk] disk status [1 Attachment]
[Attachment(s) from D. Atkinson included below] The testing and tweaking phase of the disk is about done. This year
we've had our best convergence results ever. I've attached a pdf of
stats comparisons so that you can see how well things have come
together. I'm still fixing some penalty and assist issues (a problem
every year), then I'll migrate the split disk over to the combined disk.
Some error checking, then release. The disk should be out before the end
of Sept. Thanks to everyone for your patience.