It's the back-to-backs, not the travel.

A couple years ago I built a model to project the Utah Hockey Club's first season, and I missed big. I had them pegged as one of the best home teams in modern NHL history. They turned out better on the road than at home. The exact opposite. I wrote a public post-mortem of everything I got wrong.

The thing is, I had the right instincts about what wears a team down. Travel, rest, back-to-backs, time zones. What I got wrong was the model itself. Right ingredients, bad model. So this year I didn't start over on the ideas. I fixed the model. The 2026-27 schedule just dropped, so here's what the fixed version says.

The Projection

The Mammoth land around 95 points and roughly a 3-in-5 shot at the playoffs. A low-end playoff team. Not a lock, not a lottery team. Essentially their 2025-26 selves (92 points), now stretched over 84 games. The 80% range runs from 79 to 112 points, and I'm reporting it that way on purpose. The last time I gave this franchise a single confident number, it missed by 36 points.

Home versus road splits about how you'd expect once the home edge is set honestly: roughly 51 points at home, 45 on the road. No home-ice juggernaut claim this time, because the data no longer supports one for anybody.

What I Did Differently

The old model assumed home-ice advantage was worth a fixed 50 Elo points — about a 57% edge for an even matchup. When I re-estimated it on the last three seasons, it came out closer to 29 Elo, a 54% edge. Home advantage has been quietly shrinking across the whole league, and a projection that hardcodes the old number re-creates exactly the home-ice mistake that embarrassed me the first time.

Then the part I skipped in 2024: I backtested the projection procedure itself. Before trusting a single number about 2026-27, I ran the same preseason-only method on the 2024-25 and 2025-26 seasons — frozen prior-year ratings, schedule features, no peeking — and scored it against what actually happened. Per-team error came out around 13 points, and I tuned the uncertainty bands so that 80% of teams actually land inside their 80% interval. That last sentence is doing a lot of work, and I'll be honest about it below.

What Actually Wears a Team Down

I built this to measure schedule impact, so I gave every piece its own honest test: distance traveled, rest, back-to-backs, games packed into a week, time-zone swings, cumulative jet lag.

The clean signal is rest, and it's specific. A home team on the second night of a back-to-back drops from a coin flip to about 45%. Catch the other team on one and you're up near 55%. That effect is real, it's the right size, and it's the second-strongest thing in the model after raw team quality.

Here's what surprised me: the stuff everyone blames barely shows up. A 2,000-mile trip is worth maybe a point. Time zones and jet lag come out weak and muddy, because a team that “traveled east” is usually just the road team, and you can't fully pull those apart. In 2024 I hung the whole model on time-zone fatigue. Built properly, that effect mostly evaporates. It's the back-to-backs that hurt. The flying, not so much.

And the last twist, the reason Utah's schedule is worth about half a point and not five: even the real back-to-back effect washes out over a season, because the league schedules everyone into about the same number of them. Utah's roughest stretches (a 4-in-7 in December, a 5-in-9 in January, a 6-in-12 in late February) cost real win probability on those nights, and still don't move the season. Fatigue decides nights. Over 84 games, it mostly cancels.

The Whole League

Since the model has to simulate every team to place Utah, here's the full board. Read it as team strength carried over from last spring plus schedule, not as a roster projection — more on that limit below.

If you know hockey, a couple of these will look off. Buffalo at 4th, Toronto at 28th. That's not the model being clever — it's the model carrying each team's late-2025-26 form forward and knowing nothing about the summer. Which is the whole point of the next section.

What This Model Refuses To Claim

The real lesson from the 2024 miss wasn't any single fix. It was that unstated assumptions become public numbers. So here are the ones I'm stating out loud.

• It's roster-blind. The talent rating is frozen at where each team finished last April. It knows nothing about trades, free agents, aging, or which goalie is starting. If a ranking looks wrong, this is almost always why.
• The probabilities aren't precise. With a per-team error near 13 points, I can't honestly tell you Utah is 62% and not 58% to make it. Round it to “about 3 in 5.” Anyone quoting a decimal off a model like this is selling confidence it doesn't have.
• The intervals are calibrated, not proven. I tuned the uncertainty so that 80% of teams historically land inside their 80% band. That's an honest way to size the bands, but it means the coverage number describes the tuning, not an independent test. Different thing.
• The projected spread is compressed. The best team here is ~106 points, but some team will hit 115+. That's not the model lowballing — a preseason average should be narrower than the final standings, because it can't know in advance who gets hot. It expects someone to run away with it. It just won't pretend to know who.
• The schedule effect is observational. The travel and rest numbers come from historical games, where “traveled east” still partly means “was the road team.” I report the schedule number as a model decomposition, not a proven cause. That confusion was one of the original seven failures, and it doesn't fully go away — it just gets smaller and gets labeled.