It seems like any predictive ability almost gets thrown into a blender by the impact of injures and young star breakouts / stagnations. I just did a basic "potato" comparison of just assuming teams finish in the exact same spot they did in 2021 and it seemed about as accurate in its predictive capability. Simply taking last season and assuming a repeat of 2021 was off by 5.35 spots while this model's predictions were off by 7.35 (which improves slightly to 6.7 if I remove Seattle from the equation which was the least accurate). Just for fun I even took the previous seasons rankings and shifted teams up/down by + to -12 spots randomly and got a similar distribution (off by ~6.3 spots average). I did the same with goal differential and got similar results, but don't know if extrapolating current per game goal differential over the full season is right since the numbers seemed wild (or we're on pace to have two teams with -100+ goal differential
).
That said - I imagine at the start of the season it's probably... relatively accurate? It's just stuff like Kaprizov, Husso + healthy Tarasenko, Teravainen, Barzal's slump, Whatever's going on in Chicago / New Jersey - the impact of those things are basically unknowable at the beginning of the season that makes it pretty random.
Anyways, 2c from the peanut gallery.
out of curiosity, did you use the last projections that I posted right before the season started?
“Spots” is not a great metric for this for various reasons and it’s not really what I’m going for. It’s really about the complements for me.
Understanding which components are projectable and seeing if I get the gist right in terms of what a teams strengths and weaknesses are, and then understanding what went wrong. That’s what’s most interesting to me.
as I posted before, If I was correct about seattles ranking but had all the components wrong i wouldn’t consider that a win.
but yes, obviously injuries happen and there are trades not to mention COVID this year and the islanders bizarre arena situation, lots of things will just always be unknowable.
When you do an exercise like this, you’re running the season 10,000 times and taking an average. You’ll have simulations where the Canucks got 50 points and finished in last, and you’ll have simulations where the Canucks had 130 points and ran away with the presidents trophy. But you’re just taking an average, what’s the most likely result?
in the real World of course we don’t run the season 10,000 times and take an average, so what we get is what we get. That’s why I say, even if the model is 100% perfect, it’s still not going to nail every teams ranking. Because the average result is not what we get. We get one of the 10,000 results. That’s why they play the games.
that said, with something like Seattle, I don’t think I had a single simulation where they were this bad. So that’s interesting in and of itself. And I have an idea of how to address that going forward.
