National Basketball Association

Hopefully possible win projections 2016-2017

Published Jun. 30, 2017 6:28 p.m. ET

foxsports

With the 2016-17 NBA season only a few short weeks away and the release of defending champion Andrew Johnson’s “Highly plausible” win projections a few days ago, the time has come for me to release my own “Hopefully possible” win projections.

Win projections are an exercise in repeated accuracy. Appropriately assigning player value, predicting player value for the next year, predicting minutes, and accounting for the influence of game to game characteristics on wins all need to be done accurately for the final projection to be near the eventual results. Even if all the above is done well, a major injury can play an outsized role.

For fun, I also simulated the season and resulting draft lottery 10,000 times to estimate the range of wins a team may fall into, the percent chance of netting certain draft picks, and the percent chance of breaking certain records. Those details are all listed below. I also included Andrew’s projected wins, and the current Pinnacle line. Andrew is a bit closer to the money line than me (though not by much), so that probably means he’s more right.

A few thoughts

Golden State is the only team that has a chance of breaking the wins record. It’s a sizeable chance, 21.43 percent. Adding Kevin Durant helps.

I am way higher on Oklahoma City then both Andrew and Pinnacle. Part of that might be due to differences in valuation of Russell Westbrook. A BPM-RPM blend (or similar) actually attributes more of Oklahoma City’s positive play last year to Russ over KD. PT-PM disagrees. Projected minutes could also play a role.

Boston’s future is fun. A 99.18 percent chance of making the playoffs, a win range that stretches into the high fifties, and a 98.64 percent chance of securing a top ten pick this year, courtesy of the Brooklyn swap.

Sacramento’s prediction jumped out at me as something that I didn’t expect, but Andrew and I are both significantly over Pinnacle on them. You can best believe that an age-adjusted BPM-RPM blend for Boogie at 26 is the main driver behind a higher-than-expected Sacramento projection.

If I had to pick one team that I feel like I will be wrong about, it’s Detroit. Between Andre, the market, and the coaching behind that organization (which I make no adjustment for), that projection strikes me as ripe for being wrong. Andrew is a projected under on them too.

If I had to pick a second team that I feel like I will be wrong about, it’s Milwaukee. The Middleton injury hurts them a LOT, so if new breaks that gives him a quicker recovery, they might improve markedly on this projection. A projection of +1.18 points per 100 for Giannis this season strikes me as obnoxiously low, and if there’s a player who doesn’t fit the conventional model, it might be Giannis.

Andrew and I are both way lower on Minnesota, and way higher on Brooklyn.

The Chris Bosh injury will definitely take away several Miami wins this year, and it might take away Christ Bosh’s career, which I feel really upset about.

Orlando’s Oladipo for Ibaka trade might not have the effect they hope it to have. But I’ve been very, very, very wrong before.

Philadelphia has a 99.83 percent chance of having at least one top ten pick, and a near 44 percent chance of having two. The 76ers own the Lakers’ pick provided the pick falls outside the top three. They also own swap rights with the Kings, though the odds of that swap being executed are very small. A little mentioned fact relative to the Lakers pick is that that the fourth overall pick is the most likely draft lottery outcome for the worst overall team. Another note on Philadelphia: Joel Embiid and Ben Simmons (if he plays), are the only rookies I project to have a positive impact this year. Rookies are usually bad.

In my pursuit of making accurate win projections, I attempted to mimic the statistic that Andrew has used for player value in his previous APBR-contest-winning win projections. Andrew has detailed his methodology for his player-tracking plus-minus statistic several times, but essentially, player-tracking plus-minus is an average of a regression based prediction of multi-year Regularized Adjusted Plus-Minus (RAPM) predicted using player-tracking statistics, and the multi-year RAPM it is predicting.

Out of a desire for self-reliance and in an attempt to eek out competitive advantage along the same route of innovation that Andrew has led us on, I attempted to assign player value with a far more complicated approach, creating my own multiyear RAPM, my own player-tracking regression framework, and my own prior-informed RAPM. Though the end result of this experimentation definitely passed the laugh test (for me), I found it to be not quite as predictive as currently available statistics like Real Plus-Minus (RPM) and Box Plus-Minus (BPM), and was sadly forced to not include it in my player value computations. In comprehensive out-of-sample testing I found that a multiyear weighted blend of RPM and BPM performed best at predicting wins, as it consistently produced the lowest error relative to other statistics and statistic blends I tried.

With existing players I was able to use their previous seasons’ RPM and BPM, and adjust that value for age based on the age curves of historical NBA players. For rookies, however, I had to take another approach. To that end, I estimated rookie value in two separate approaches. First, I took the result of my draft projections, which predict a player’s two year peak, and appropriately scaled that into my blend range (as my draft projections were performed on a “wins-added” scale), and used age curves to reduce that draft projection value from a career peak to their upcoming rookie year. Next, I created a model that predicted a player’s first year BPM, based on attributes like first year age, height, weight, and draft position. My final result for rookie value was simply the average of these two approaches.

To predict the amount of minutes each player was going to play, I created a random forest model that used the amount of minutes a player had played over the last two years (or one year), as well as variables and interactions concerning previous years’ BPM, height, age, weight, experience, and draft position. The rookie model followed the same logic, but only used age, physical characteristics, and draft position to estimate total minutes played. Next I manually adjust for known injuries with best guestimates of how it will influence their total minutes played, injuries like but not limited to, the injuries suffered by Chris Bosh, Khris Middleton, and Ben Simmons. When the total of predicted team minutes are undoubtedly not exactly the same from team to team, I adjusted each player’s minutes based on their predicted percentage share of the team’s, and either give much more or took much less of the minutes realignment to the predicted top ten players of the rotation.

The sum of the product of my player value RPM-BPM blend and the minutes a player had played were then used to estimate the strength of each team, and the difference of that team strength was used to help train predictive model on historical NBA data at the game level. Also included among the predictors of those models was the altitude of the game, the rest each team had had prior to the game, and interaction variables that accounted for how long the game was, and whether or not that meant a benefit for the better team or the worse team. More possessions decreases variance, which means that a bad team can win more games over the course of a season simply by playing slower, by virtue of giving the better team fewer opportunities to be better. I tested several different modeling methodologies for the task of predicting games, but found that an ensemble of logistic regression and gradient boosted regression predictions performed the best out of sample.

To arrive at total wins, I simply aggregated the estimated probabilities and rounded to the nearest win. For fun, I also simulated the season and resulting draft lottery 10,000 times to estimate the range of wins a team may fall into, the percent chance of netting certain draft picks, and the percent chance of breaking certain records. Those details are all listed below. I also included Andrew’s projected wins, and the current Pinnacle line. Andrew is a bit closer to the money line than me (though not by much), so that probably means he’s more right.