The Clock, Part 2: What we found

That's the one I get asked, somewhere between the second beer and the check, and I wanted to come back with an answer that would embarrass every C- on the internet. So I built the thing to settle it: every roster in the league rebuilt from the box scores, every player scored four different ways, the season run 200,000 times with and without each move Minnesota made.

At 1:40 in the morning it came back and said worse.

I ran it again. Worse. I checked the roster, the contracts, the injuries. Worse. Not by a lot, but every version of it, and the graders had just gotten a free win from my own model.

So I had to find out why. The why turned out to be a twenty-one-year-old I had never thought about for more than a minute.

The kid from Utah

When the Green trade came across my phone that Saturday in August, I didn't give Cody Williams much thought. A throw-in. A long wing who could soak up some minutes in January when the injuries come, because they always come. I still think that's roughly what the front office was thinking.

Then I ran the season, and the model came back with something more interesting than that.

Start with the roster, not the cap. Out: Randle, Reid, Conley, Anderson, Donte on crutches. In: LaMelo, Kuminga, Cody Williams. Run the roster that finished last season against the one that will start this one, and the first time, every version of the model said worse. Not by a lot. All of them.

Every version, because I don't trust any one of them alone. I score every player four ways. Two of the four are close to twins, so the headline number averages three of them, and nothing gets called a finding unless it holds in all four.

So I went looking for the reason the way you look for your keys, everywhere obvious first. It wasn't Randle. It wasn't Naz. It wasn't LaMelo, who the model likes more than the grades do. It was the throw-in.

He was the No. 10 pick two summers ago, a long wing with a jumper the scouts kept promising, and he came to Minnesota on the last Saturday of August as the piece Utah put into the Green dump, the throw-in that made the arithmetic work. The numbers from his two Utah seasons are rough. The model scores him at -4.39 points per hundred possessions, lower than every player in the league's projected rotations but one. The front office says it sees untapped potential, which is what you say about a former lottery pick who arrived in a salary dump; the honest version is that they took him because he's young and long and the alternative was nothing.

His minutes. The first runs gave him 16 to 19 a night, because the model did what a lot of people do: it saw a guy who played 24 minutes for Utah and assumed he'd play real minutes here. But a role on a bottom-ten team isn't a role. It's a vacancy. Once I stopped the model counting Jazz minutes as Wolves minutes, for every player in the league who changed teams, Williams fell out of the rotation entirely, the 12th or 13th man depending on how you order it. A roster that got better almost everywhere else had been dragged under by one player's minutes.

With him on the bench and everybody healthy, the summer helped.

That should have been the end of it.

It wasn't. Then I built the thing I should have built first: people get hurt. Every player now carries a chance of missing games, drawn from players like him, and the season gets played night by night with whoever is left. On a healthy night Williams sits. But healthy nights don't last, and across the whole season the model hands him about 20 minutes a team game, nearly all of it on nights somebody above him is out. Last season's deep-bench players in his spot, the ones who'd had a real role somewhere else, the way he did in Utah, got 14.7. So the model probably plays him about 5 minutes too much. In the model, those minutes cost Minnesota about a win.

The beat reporting has him outside the rotation barring an injury. In the preseason, with every regular available, he went in after the two-way players at Milwaukee, first off the bench after the regulars at Indiana, where he played about 18 minutes, and not until the fourth quarter against Detroit, for the last 8 minutes. With regulars sitting, he played about 19 at Miami and started at New York in Kuminga's place, about 15 minutes, fewer than Dosunmu got off the bench. The headline number agrees with the reporting about the healthy nights; it's the injuries that bring him back.

Then there's the fork. Give him real minutes and they come out of players the model trusts more. Before the age adjustment, every version says the summer hurt once he's above 9.2 a night. With it, even at 16 a night, two of the four still say it helped. That's Finch choosing him with everybody healthy, not injuries handing him the minutes. So the number I'll watch at game twenty is simpler: his minutes per team game, every minute counted and a game he sits counted as zero, against the midpoint between the model and last season's deep bench. That's 17.3. Above it, the model has his minutes right. Below it, Minnesota is a little better than I'm telling you.

What the fork is really saying

Here's the summer in one sentence: the gain is biggest on the nights they're whole, and it shrinks every time somebody below the line has to play.

The model isn't making a prediction about Cody Williams. It's making a statement about the roster. The gap between the eight guys who are supposed to play and the guys behind them is so big that the moment one of the guys behind them has to play real minutes, the summer stops being a clear plus. And that happens most nights: last season the typical team had the eight players it used most all play in only 25% of its games, and even its top five in only 58%.

Six players carry this team on the model's books. Beringer is 19. The model doesn't believe in Terrence Shannon the way the coaching staff does (and so do I, but that's a different article), a call game twenty will grade, and if the staff is right the floor is higher than I'm telling you. But the shape doesn't change. The summer traded a nine-deep team for a six-deep team with a better top.

So this is a team that has to stay healthy to be what it wants to be.

Which left the question I had been walking around all summer. Whose minutes were those supposed to be?

Weather

Game 4 against Denver, in a series Minnesota won. Donte goes down, and you know from the way he goes down. He'll miss most or all of this season, and the front office didn't bring in an established wing to cover for him. You can count that as a decision, since they could have, or you can count it as weather, since nobody plans for a torn Achilles in April.

The model says it costs 0.33 points of title odds on its own. Count it against the summer and the model splits on the whole. Count it as weather and, once the model adjusts for age, every version says the summer helped.

Those aren't two separate problems. A healthy Donte is exactly the player who takes Cody Williams's minutes. The injury and the kid's role are the same wound viewed from two sides, and every argument about whether the summer was good or bad is secretly an argument about who plays the wing minutes in November.

So my honest answer to the bar question: better, if they stay healthy. Nobody stays healthy all season, so ask me at game twenty how it's going.

What I can give you before game twenty is the part of the model I'd bet on. But first you should know about the night I couldn't.

The honest scoreboard

A model that's never been wrong in public is a model nobody has checked. Here's where to check this one.

The model had the Hornets at 6.39% to win the title. Charlotte. A team the market prices at 0.81%. I stared at that number for a long time, and I'll admit my first instinct was to look for a reason it might be right. What I found instead was a bug. The pipeline every player rating is built on was crediting some baskets to the wrong team, so I fixed it at the source and reran the whole thing. Minnesota's number barely moved, which is the part that let me sleep.

The bug was real and the fix was right, and it didn't solve Charlotte. The model still has the Hornets at 5.75% to win the title, 48 wins against a Vegas line of 39.5. This time there's no bug to find.

So I did the thing I should have done in August. I rebuilt the model as of last October, with only what it would have known then, and asked it about last season. It missed last season's teams by about 4 points of net rating each. The market missed by about 3.5. When the two disagree, the market has the better record, and I'm not going to pretend otherwise. What failed that test got fixed. Every fix was checked against last season before it went in. One fix that looked right and made the test worse got dropped. Then I froze the model, and everything in this series is the frozen version.

Minnesota. The model has the Wolves at 1.79% to win the title. Its four versions rank them anywhere from 10th to 17th. Without the age adjustment it's 1.38%. The market says 3.16% and sixth in the league. In wins, the model says about 42. Vegas says 48.5. If the Wolves' net rating after thirty games is above +1.8, halfway between the two, Vegas was closer than I was.

Last season the Wolves were healthier than anyone expected, and the model takes most of that luck back. The bench behind the top six is thin, so absences cost them more than they cost most teams. And the model treats a healthy scratch or a G League game like an injury, which is part of why it plays Cody Williams too much, by about a win. I tested a version that doesn't treat a healthy scratch that way. It called last season no better, so it didn't go in.

Charlotte isn't the only one. The model has the Pacers at 32 wins against a line of 43.5, piling Haliburton's first season back from an Achilles on top of a rotation it expects to miss a lot of games. Somebody is wrong, and we'll know by December. The rule is the same for every team: after thirty games, whichever expectation a team's net rating is closer to, the model's or Vegas's, wins. For Charlotte the halfway line is +1.2; below it, Vegas was closer. For Indiana it's -1.4; above it, Vegas was closer. If Vegas wins those, the model owes you an explanation.

If you want to know whether to trust the model, don't watch Minnesota. Watch Charlotte.

What I'm sure about

Now the part I'd bet on. The sure things are why I'm not worried the way the grades are.

Kuminga first. The question isn't whether Jonathan Kuminga is a great player. The question is whether he's better than whoever else would have played the four, and without Kuminga the model's rotation gives those minutes to Cody Williams. The model leans yes: better by 0.29 points of title odds, and on the right side in every test, but not by enough in every version to call it settled. It leans yes. It isn't proven. Put Trey Lyles at the four instead and the lean goes the other way, toward Lyles, though not by enough to call that either. The one that isn't close is Beringer. If Joan Beringer takes those minutes, the model would rather have Beringer, by a lot, in every version. It says that on the strength of 7.9 minutes a game of rookie evidence, more than a third of it in garbage time. Which is either the most interesting fact on this roster or a small sample doing what small samples do. Probably both. To be completely honest, I am very high on Beringer, and would love to see him take a huge leap this year, but he is very young. So, we'll see.

The trade itself is cleaner. LaMelo Ball in is the one move on the ledger that is clearly positive: 0.71 points of title odds, and every version of the model agrees. Naz Reid out is clearly negative: 0.28 points the other way, and again every version agrees. Randle out doesn't get a sign, because the model's four versions can't agree on whether he was helping. The Dosunmu re-sign doesn't get a sign for the same reason, and what that contract really cost is a cap story for Part 4. So if you want the trade in one sentence: they got the point guard, and Naz, not the picks, was the real basketball price. I've made my peace with the picks. I have not made my peace with Naz, and this time the numbers are on my side: losing him is a cost in every test I ran. Every verdict's four cells are on the methods page.

And then there's the one I can't be sure about and can't stop thinking about, because it isn't a number from this summer. It's a look on a face from May.

The May I keep coming back to

If you watched the San Antonio series you already know the shape of it. The Spurs put one defender on Ant, ran him off the three-point line, and made him make plays out of the space between the free-throw line and the rim. He had good nights. He also had nights where the whole offense was him trying to beat a scheme by himself, because there was nobody else on the floor the defense had to account for as a creator, and you could see it on his face by the fourth quarter. I remember thinking it looked like a problem a roster causes, not a problem a player causes. I didn't have a way to check that in May. I do now.

The finding: assign one good defender to Anthony Edwards and he loses more than almost any scorer in the league. For most stars, a good defender costs them a little. For Ant it costs more than for almost any of them, and over three seasons more than for any of them. Take the 150 highest-volume scorers from last season. Edwards sits at the 1st percentile: almost nobody in the league loses more to his primary defender than he does. Widen it to three seasons and 245 scorers and he is the lowest of all of them, the single most defender-suppressed scorer in the sample. Twenty-one different defenders drew the assignment last season, so this isn't one bad matchup; it's what happens whoever the assignment is. That's not a stat about Ant's talent. It's a stat about what a defense can do to a player, and the simplest reading is that there was nobody else on the floor who could create a shot.

Which is what the rest of the numbers say too. Last season 61% of Ant's makes were unassisted. LaMelo, in Charlotte, was at 55%. The league median is 35%. Minnesota just put two of the most self-created scorers in basketball in the same backcourt, and the obvious worry is that they'll get in each other's way. I had that worry. The base rate doesn't. When a high-usage player gains a new high-usage teammate, the player who was already there loses usage and doesn't lose efficiency. At the star level that's 1.5 points of usage given up. True shooting actually went up, by 0.8 points, across nine cases. Fewer shots, not worse ones. And a starting five with Kuminga in it has a combined usage at the 96th percentile of the starting fives we could measure. There were 7,380 of them, which is the polite way of saying somebody is going to give up shots. The base rate says that's fine. Somebody always does.

That's why Ball is here. Not because the front office fell in love with a highlight reel. Because the single most measurable weakness on the roster was that a defense could take Anthony Edwards out of a series by assigning one good player to him, and the fix for that is a second creator. Whether LaMelo is the right second creator is a fair question. That the job needed filling isn't. The trade was an answer to the thing I watched happen in May.

The arguments we have in September

Most of the things fans argue about in September, the things I argue about, can't be measured. We tried, and I went in wanting to prove the podcasts right. That doesn't mean size and pace and matchups don't exist, or that coaches are wasting their time. It means that once you know how good two teams are, knowing their style doesn't help you predict who wins, at least not with the tools I have and the seasons we've got. That's a statement about prediction, not about basketball. But it's a statement the odds are quietly making too, and it turns out to be right.

The style model is the one everybody wants: pace, three-point rate, rim rate, offensive rebounding, turnovers forced, size, the stuff you'd put on a whiteboard. Adding it to the predictions made them worse. Not much worse. Worse.

Eight regular-season traits are supposed to translate to the playoffs, the ones you hear every May. I tested them on 195 playoff series going back to 2014. Zero of the eight predicted who won a series once you knew how good the teams were. Not the three-point rate. Not the size. Not the pace. Not the thing your favorite podcast is sure about.

Three more, because they're the ones I hear most.

Of the eight, "defense travels" is the one everybody's surest about, and it gets no support here. If anything the data leans the other way, not by enough to call it, but not toward the cliché either.

Versatility, the idea that some teams match up well with everybody, turns out to be a fancy way of saying they're good. The better team is the better team against everybody.

And the "who guards Ant" arguments come out the same way one level down, with one exception, and you can guess who. Of 77 real matchups across the West's contenders, a scorer and his primary defender, two went the scorer's way by more than usual, which is about what chance alone gives you. Four went the defender's way by more than usual, against less than one by chance, and three of those four were defenders guarding Anthony Edwards. Matchups mostly don't matter. A good defender on Ant does, whoever he is, which is the May finding again.

So when someone tells you the Wolves will be fine because they're big, or doomed because they're small, ask them which of the eight it is. I've started doing this. It has not made me more popular.

Health, and the path

Health. Take Ant or LaMelo out for the playoffs and about two thirds of the title odds go with him: 68% for Ant, 62% for LaMelo. Take Rudy out and it's 41%. Oklahoma City and San Antonio each have one player who takes half or more. Minnesota has two. It's the bench story from the other end: this team needs to be whole more than most. Part 4 is where that gets its due.

The path. The most likely seed is 7th, which means the most likely version of April is a play-in game. It lands there 24% of the time. Top six, 46%. If they make the field, San Antonio or Oklahoma City is the first-round opponent 47% of the time, still the biggest single problem on the calendar. They reach the second round 28% of the time. From there, having escaped the first round, they win it all 6.4% of the time. The mountain is still the first round, and nearly half the time it has Wembanyama or Shai standing on it.

One thing the model does in the playoffs that it doesn't do in the winter: it plays the top of each rotation more, the way coaches do. That helps Oklahoma City, San Antonio and Denver more than it helps Minnesota, which would be at about 2.4% to win the title without it. That part is checked against how playoff minutes actually get handed out, not against who won.

So I got curious. If the odds can't see what wins, what did the last eleven champions actually have? That's Part 3.