first 20 games tell

Why Your First 20 Games Tell You More Than You Think — And Less Than You Fear 

Twenty games into the 2025-26 season, our model projected Carolina to finish with 114.8 points. They finished with 113. Twenty games into the same season, our model projected Buffalo to finish with 74 points. They finished with 109 — a 35-point miss.

Same sample size. Same method. Wildly different results. Both numbers are instructive, and understanding why they diverge is more useful than either number on its own.

The conventional wisdom — and why it's incomplete

The standard line in hockey is that early standings are noise. Don't overreact to a hot or cold start. Wait for the sample to grow. It's not bad advice, but it's incomplete, and incomplete advice is often more dangerous than no advice at all — it gives decision-makers false confidence in when to wait.

We ran the numbers across 160 team-seasons — all 32 NHL franchises over five seasons — to find out whether there's an actual inflection point where early data stops being noise and starts being signal. There is. It's around 20 games. Not perfectly reliable at that mark, but reliably enough to act on, with an average error of about 8.5 points league-wide.

The more useful finding isn't the number 20. It's what determines whether your team's 20-game projection belongs in the reliable group or the Buffalo group.

What separates a Carolina from a Buffalo

Carolina's 20-game projection worked because Carolina is a stable system. Over five seasons, their CF% — Corsi For percentage, the share of all shot attempts a team generates at 5-on-5 out of the total attempts in the game — has a standard deviation of just 1.49. That's the lowest variance of any team in the league. When a team's underlying shot-attempt profile barely moves from month to month or season to season, 20 games of it looks a lot like 60 games of it.

Buffalo's miss wasn't a fluke of small samples — it was a signal that the organization was in genuine transition. Teams that are improving or declining don't have a stable baseline for a small sample to reflect. Their early games capture a snapshot of a system that's still changing shape. That instability is not a flaw in the projection — it's the projection correctly telling you something is shifting.

This is the same pattern we saw when we ran a similar test on Utah's 2025-26 season. Their 10-game projection was wildly optimistic (131 points) after a hot start, but it collapsed back toward reality by games 16–20, settling into the 92–96 point range — within a few points of their actual 92-point finish. The lesson: a hot or cold start can distort a small sample badly, and the way to catch it isn't to ignore the data, it's to check what's driving it.

PDO: the noise filter

That's where PDO comes in — a team's shooting percentage plus its save percentage, expressed on a scale that averages 1000 across the league. A team running well above 1000 is scoring and saving at a rate that its underlying shot quality doesn't fully explain. That's usually puck luck, and it tends to fade. A team running well below 1000 is often getting unlucky in ways that will correct themselves too.

Any team with an extreme PDO in its first 20 games — hot or cold — is a poor candidate for a straightforward projection, regardless of what its record says. The record is real. The paint that produced the record won't hold.

Three questions at game 20

If you're an analytics staffer trying to give a straight answer to a GM or coach who wants to know whether the first 20 games mean anything, here's a usable framework:

1. Is our CF% consistent with our recent history? If it's in line with the last two or three seasons, the sample is more trustworthy.

2. Is our PDO above or below 1000? An extreme reading in either direction is a warning to discount the record itself.

3. Does our points pace match our underlying shot metrics? If the team is out-attempting opponents and winning, that's aligned signal. If it's winning while getting out-attempted, something unsustainable is propping up the record.

If all three line up, trust the projection. If they diverge, the divergence itself is the finding — it's telling you to investigate before anyone makes a personnel or coaching decision based on 20 games of standings alone.

The broader point

Analytics tools earn their keep not when they confirm what everyone in the room can already see on the scoreboard, but when they surface the gap between what the record shows and what's actually happening underneath it. Twenty games is neither too early to know anything, nor late enough to know everything. It's exactly enough data to know which question to ask next.

© 2026 ORRO AI Genius, LLC. All rights reserved.