No But Seriously… The Warriors Might Break The Record

Over the past month, the Warriors have gone from championship favorites in a top heavy West to a team so blitzingly awesome that when people ask me about them I just respond by shrugging and saying they’ve solved basketball. It seems almost true at this point. A few weeks ago there was an open debate on who the best player in the league was. I’m not sure anyone is having that conversation anymore.

My preseason win projections put the Warriors at a gentleman’s sixty. In the light of their 20-0 start and Houston’s struggles, those projections now look downright silly. I recently took it upon myself to revisit the modeling framework behind these projections to not only make significant improvements on these projections and account for in-season statistics, but also to come up with a percentage chance that the Warriors break the record for most wins in a regular season.

And at this point in the season, my estimate of the Warriors winning 73 or more games is:

75.32%.

That’s right. It’s not 20% anymore, or 3%. It’s 75.32%.

Of course, that percentage assumes the Warriors keep up the torrent of soul-crushing, merciless dominance that they’ve played with so far. It also assumes no major injuries.

The Warriors have been lucky enough already. They’re probably going to regress to the mean to some degree. I don’t know a way to responsibly  determine to what level that might happen, and I have no crystal ball telling me when someone will tweak their ankle, so I came up with the following simple method: I substituted the Warriors current 15.4 team efficiency differential in the team component of my win projection model (more details below) for the gradually decreasing efficiency differential that would reflect a regression to the previous season’s 10.2 efficiency differential, and reflected a similar change to the in-season player-level components of my modeling as well. This will credit the Warriors for their hot start, and will also represent the Warriors and the Warriors players’ playing the next 62 games near last season’s level. In this situation, they average winning about 5 games less than the 74 games at their current pace. They would still have a 50.33% chance of winning 70 or more games, and a 14.30% chance of winning 73 or more.

In reality, the chance of the Warriors reaching these milestones is somewhere in between the numbers in the two columns displayed in this table. Either way, it looks like they have pretty good chances of breaking seventy. Those last three potentially record breaking wins is really where the question lies.

I hope you didn’t think the Warriors were going to go undefeated. Sadly, that’s not going to happen, no matter what. NBA results are simply not that consistent. In fact, as dominant as the Warriors have been so far, pythagreon wins dictate that they should have lost twice already, (much like the record-breaking Chicago Bulls had 20 games in), and this is with double-digit RPM from Stephen Curry, a true top ten player in Draymond Green, and the so called “lineup of death” maintaining these insane statistics over more than a game of time together.

How did I arrive at these numbers? The model that arrives at this estimate considers a wide range of information. The first major component I use to predict games here is a projected year end RAPM model. The model is a neural network and linear regression blend that uses preseason information (and only preseason information) to predict a player’s season-ending RAPM. The RAPM being regressed onto here is non-prior-informed, and this first number considers publicly-available statistics like RAPM, Win Shares, BPM, and Age to make its predictions. The next major component of the modeling is an in-season year end RAPM model. Using a player’s pace-adjusted day-to-day box score statistics, a neural network model predicts a player’s season ending RAPM. If Curry drops 80 points on 75% shooting tonight, this second component will grow in value for the Warriors’ next game. After some minor tweaks that include taking some minutes away from the bottom of the roster in order to ensure minutes for each team equal 240, I create weighted sums of these first two components for each team, weighted by the tweaked minutes per game a player has played so far this season. The third component is a measure of team strength, the efficiency differential of each team going into a game, regardless of injury or who plays. With these three main components for each team, in addition to game information like days of rest and altitude, I use a logistic regression model trained on games back to 1990 to predict the outcome of each game. This modeling and variable configuration performed the best out of sample of all the modeling configurations I tried with the data I have. It correctly picks games out of sample at just about 70% and correlates strongly with game margin and Vegas money-lines as well.

If you are one of the five people who may be interested in my win projections, I would encourage you to check out this Shiny App I built for displaying them. This app updates on my end every night, and will show you my estimated chances of each team winning on the current day, using the modeling configuration I just detailed briefly. You can also adjust minutes as you see fit and see the effect missing certain players will have on a team’s winning chances, making it a useful tool to employ when a player gets injured at the last minute. (Though keep in mind there is a team component here as well, which will cause the percentage chance to underestimate the impact of missing key players). Let me know what you think!

The Warriors are quite unlike anything we’ve seen in the NBA to date. It’s fitting that they are more likely than not to win seventy games, and have a more than decent chance of beating the all-time record. We are truly privileged to be witnessing this. It’s goink really well.