NEW PERK: 2020-22 WAR Cards (Patreon)

You guys have been asking for them, and they're finally here: 2020-21 vs. 2020-22 cards are ready to go. These will allow you to gage how players' performance has changed in the early days of the new season compared to last year.

If your last name starts with the letters A-H, access them here: Click Here 

If your last name starts with the letters I-P, access them here: Click Here 

If your last name starts with the letters Q-Z, access them here: Click Here 

These cards display two sets of data:

• A player's Wins Above Replacement per 82 games percentile ranking in each category from the 2020-21 NHL season
• The same player's WAR per 82 games percentile ranking in each category from a combined sample of all games from the 2020-21 and 2021-22 season so far. 

Essentially, the numbers on the left represent the model's evaluation of the player based on the information it gets from last season, while on the right it's treating both samples as one broad data set. 

On the right side of the visualization you see the difference in each category between the first sample and the second. In most cases, these will be relatively small, but for some players these can be very dramatic. That's because of the large amount of new information the model has to learn from including many many new combinations of players and new circumstances.

An explanation for why this is the best way of incorporating new early data is below.

What Has Changed, and Why 2020-22?

The mechanism behind WAR is called RAPM, which basically absorbs all of a player's data, his context, his teammates' data, his opponents' data, etc. to form an estimate of how he is individually impacting things on the ice. The larger a sample you have, the more the model has to "learn" from and the more confident you can be in the results. As a result, you want as large a sample as possible.

So that presents a problem in the early going of the season, because there is a tiny sample - things obviously change over time and we want to keep updated on player performance without having to wait several months. But if you were looking at WAR or RAPM data from the 2021-22 season only, it would be unreliable to the point of meaningless because it has barely any data to go off of. One option would be to just shrug our shoulders and publish that early data with a "small sample" caveat, but that's just asking for really bad analysis (like the "Holl for Norris" thing from last January) and horribly misleading information. We're not interested in putting out bad data.

Patrick and I have come up with a solution that I think works best. Last season was, of course, shortened, meaning every player played a lot fewer than 82 games. So we've decided to combine the 2020-21 season and the 2021-22 season until the time comes (a few months into the season) when the 2021-22 data can stand on its own. This will give the best assessment of a player's current performance level, the best projection of their abilities moving forward, and a much better and more usable model. The WAR model will learn new information about these players every night, especially ones who have joined new teams and have whole new sets of linemates.

You might be tempted to say "but I want to see how players are performing this season, not in the past two seasons combined." But the unfortunate fact of the matter is that this early in the season, no metric exists that can tell you that. This is the best option.

So on these player cards, the 2020-21 data has been replaced with combined 2020-22 data, with no changes in weighting - this is beneficial since 2020-21 was already a small sample to begin with.

So next up is the website. I've been back and forth with Wordpress on a few technical bugs but they seem to be mostly worked out. It will be coming soon and will host a LOT of new visualizations for you guys.

Enjoy!

Jack