Findings

Six things the data says, measured on all 9 AHL seasons in this database

Pull the goalie about two minutes earlier

A goal down, AHL benches pull with 1:04 left. The model says 3:21 — and two goals down, 8:33 rather than 2:20.

3:21when to pull, a goal downbenches go at 1:04
+3.8standings points left on the table2026-27, across the league
10,955team-minutes of six-on-five measuredand 3,282 empty-net goals behind the rates

Win probability gained over never pulling, in percentage points · filled dot = model optimum, hollow = league median

Every curve is flat across its top, so the model is not asking anyone to hit 3:21 on the nose. It is saying the league median of 1:04 sits on the wrong side of the peak.

Every deficit in detail

Down 1

3:21

left in regulation

Benches pull at: 1:04 left
Win probability, never pulling: 23.2%
…pulling when they do: 24.4% (+1.2 pts)
…pulling at the optimum: 25.2% (+2.1 pts)

Down 2

8:33

left in regulation

Benches pull at: 2:20 left
Win probability, never pulling: 7.4%
…pulling when they do: 8.4% (+1.0 pts)
…pulling at the optimum: 9.4% (+2.0 pts)

Down 3

20:00+

before the third period starts

Benches pull at: 1:20 left
Win probability, never pulling: 1.9%
…pulling when they do: 2.3% (+0.5 pts)
…pulling at the optimum: 3.3% (+1.4 pts)

The rates the model runs on
Deficit	Goalie in net				Goalie pulled				6v5 min
	GF/60	GA/60	xGF/60	Shots/60	GF/60	GA/60	xGF/60	Shots/60
Down 1	2.61	2.23	2.71	30.4	9.57	29.46	7.69	70.7	2,878
Down 2	2.61	2.60	2.79	31.3	6.92	19.07	6.04	56.9	4,558
Down 3	2.78	2.92	2.82	31.6	4.80	7.16	4.06	38.7	3,518

A goal down, the extra attacker buys 3.8 times the scoring rate at 13.9 times the rate against. A terrible exchange on its face, and still worth making: a goal against when you are already losing costs far less than the goal you need is worth. That asymmetry is the whole reason the trade pays.

A 2026-27 season replayed this way

Games with the decision

116

someone was behind late enough to pull

Extra wins

+1.4

one more per 82 such games

Where the points come from
Deficit	Games	Pulled earlier by	Points gained
Down 1	41	2:00	+2.0
Down 2	39	5:40	+1.0
Down 3+	36	18:40	+0.7

Most to gain
Club	Games	Points taken	Gain
IA	7	3	+0.26
PRO	6	0	+0.24
HAM	6	0	+0.23
CGY	5	0	+0.22
HFD	6	0	+0.21
TEX	5	2	+0.19
TOR	4	1	+0.15
BEL	4	0	+0.13

No game is simulated from the opening faceoff. Each is replayed only from the first moment the optimal rule and current practice disagree — real score, real clock — and only that first moment counts, because after it the two worlds have separated. The replay maximises standings points rather than win probability, which pushes the thresholds a little later, to 3:04 and 8:00: reaching overtime banks a point.

How this is measured, and where it is weakest

Does the model know what it is talking about? Run it over every season on file under the policy benches actually use and it expects 1139 standings points from those 8,507 situations. The teams took 1,125 — a miss of 1.2%. The counterfactual is only worth reading because the factual lands.

The feed never records the goalie leaving the net, so six-on-five ice time is inferred. A team with an empty net cannot concede an ordinary goal, so the rate at which ordinary goals against go missing — against a baseline from the same teams minutes earlier — measures how many nets are empty. Two checks on it: goaltender time on ice, an entirely separate column of the box score, implies about the same 11,000 team-minutes without being used here; and the goals this says were conceded with a goalie in net average thirty feet of shot distance, exactly like ordinary goals.

Down three is the weak case. By three goals the leading team has stopped hunting the empty net — it scores into it at 7 per sixty rather than 30 — which is what pushes the answer past the edge of the period. Read that one as “far earlier than anyone does” rather than as a time. Hold the leading team to its up-two rate and the model says 5:19; to its up-one rate, 3:19.

Pull the goalie about two minutes earlier

Every bench in the league waits too long, and it costs them standings points.

A lead flatters the club holding it

Clubs in front outscore their opponents 53.5% of the time. Take empty-net goals out of both sides and it is 48.1% — at even play the club protecting a lead is being outscored.

48.1%goal share with a lead, empty net removed53.5% with those goals counted
5.4points of goal share the empty net is worththe entire margin, and then some
5.35 → 5.58goals per 60, tied then a goal aheadthe game does not slow down when someone leads

League, 2026-27	GF–GA	GF%	GF% without empty-net goals
With a lead	289–251	53.5%	48.1%
From behind	251–289	46.5%	51.9%

Why a lower goal share can still be the right trade

A leading club is not trying to win the next ten minutes, it is trying to run out of them, so trading goal share for a slower game can be correct. It is finding 01 in reverse: the trailing club wants variance badly enough to hand over an empty net, and the leading club should want to kill variance for the same reason. A shell is the posture that goes with it — collapse into a 1-3-1 or a box around the slot, concede the outside, kill anything from the middle.

Holding a one-goal lead, twenty minutes to go
How the leading club plays	Source	Goals/60, both	Its goal share	P(win)
What AHL leaders do	measured	5.58	46.8%	.744
Like a tied game	measured at level	5.36	50.0%	.774
A shell: scoring −40%	hypothetical	3.35	42.1%	.783
A deeper shell: −60%	hypothetical	2.23	39.9%	.829

The hypothetical shells win, so the logic is sound. The trouble is that the game does not slow down: to justify a 46.8% share, total scoring would have to fall about 18%. It rises 4%.

The honest limit on this

Those rates are an equilibrium — the leader sitting back and the trailer pushing — so part of the extra scoring is the trailing club’s doing rather than a failure by the leader. Nothing here observes what would happen if a leading club played it differently, so read the table as the price of the trade, not as proof that any particular bench is getting it wrong.

What is safe to take from it. Judge a club that spends the season in front on its even-play numbers rather than its goal share, and the reverse for one that spends it chasing — that part needs no counterfactual. Regulation only; overtime is three-on-three and always starts level. Per-club versions are on each club’s page.

The third period is not a better shooting period

Shooting percentage climbs 9.2% → 10.0% → 11.0% across a game, which looks like tiring legs and desperate hockey. Remove goals scored into an unguarded net and the third period falls back to 9.4% — below the second.

11.0%third-period shooting, as a box score counts itthe figure that makes the story
9.4%the same period, empty-net goals removed10.0% in the second, 9.2% in the first
3,094third-period goals scored into an empty net15.7% of every goal in the period

Regular season, all 9 seasons on file
Period	Shots	Goals	Shooting %	Empty-net goals	Shooting % without them	Goals ÷ expected
First	184,491	16,977	9.2%	2	9.2%	0.997
Second	192,547	19,245	10.0%	4	10.0%	1.009
Third	178,867	19,664	11.0%	3,094	9.4%	0.998

The last column is the one that settles it. Goals land within a percent of what the chances were worth in all three periods — nobody finishes better late, and nobody finishes worse early. What changes is the quality of the chances, and in the third period, whether there is a goaltender in the net at all.

The one period effect that is real

The second period is genuinely the best one to be shooting in — 10.0% against 9.2% in the first — and it is chance quality rather than finishing. Expected goals per shot peak there too (0.0991 against 0.0923 and 0.0944), and there are more shots as well: 192,543 against 184,489 and 175,773.

That is the long change. For twenty minutes both benches are at the far end from their own zone, tired defenders cannot get off, and the odd-man rushes that follow are worth more than the shots a rested team gives up. It is the one thing in this finding that survives taking the empty net out, and it is worth about eight-tenths of a point of shooting percentage.

What counts as an empty-net goal here

A team with its net empty cannot concede an ordinary goal, so the feed marks every goal against it en_against. That label is what is removed — 3,100 goals, all but 6 of them in the third period. The expected-goals model excludes the same shots from its fit and scores them at zero, since beating an unguarded net is not a thing an average goaltender can be measured against.

What is safe to take from it. Any per-period rate you read anywhere — shooting percentage, save percentage, PDO by period — is inflated in the third unless it says it took these out. This is the same distortion as finding 02 above, arriving from a different direction: the empty net is the largest single contaminant in AHL scoring data.

The third period is not a better shooting period

Every goal-scoring rate rises through a game. Take the empty net out and the rise disappears.

The second shot is the best shot in hockey

An attempt taken within four seconds of the club’s own previous one goes in 22.9% of the time. The first attempt of a possession goes in 8.7% of the time. Everything a team does to generate one more shot at a loose puck is worth 2.6 ordinary chances.

8.7%first attempt of a possession524,352 of them
22.9%second attempt, inside four seconds33,198 of them
14%of all goals come off that second attempton 6% of the attempts

Attempts by their place in a chain, all 9 seasons
Attempt	Shots	Goals	Conversion	Expected
First of a possession	524,352	45,614	8.7%	8.6%
Second, within 4s	33,198	7,602	22.9%	22.8%
Third	3,301	855	25.9%	23.9%
Fourth or later	593	87	14.7%	17.1%

What to do about it. Chances are not interchangeable, and a club’s shot total says nothing about how many of its attempts were the good kind. Every club page carries its scramble share for that reason: two clubs on thirty shots a night are not playing the same game if one of them gets to the rebound.

How a chain is defined, and why the expected column matches

Consecutive attempts by the same club within four seconds are one chain, guarded by period as well as by game — 19:58 of the first and 0:02 of the second are four seconds apart on the clock and an intermission apart on the ice. The AHL feed charts no faceoffs and no zone entries, so possession has to be inferred from the shot clock this way; there is no way to know whether the puck actually left the zone in between.

The expected column tracks the actual one almost exactly (8.6% against 8.7% on first attempts, 22.8% against 22.9% on second). That is a consistency check rather than a discovery: the model is fed the gap since the club’s own last attempt as a feature, so it is supposed to know. What it shows is that the site’s expected goals are not a location model wearing a fancier name — take second chances out of it and it would misprice a fifth of all goals.

Fourth attempts and later are the one row not to lean on: there are only 593 of them, and at that depth the puck is as likely to be smothered as shot.

The second shot is the best shot in hockey

An attempt inside four seconds of the last one converts at two and a half times the rate.

How the puck arrived beats where it was shot from

Hold the location fixed — inside forty feet, in the middle of the ice — and the same shot is worth 15.9% on a fresh possession and 12.9% off sustained pressure. A shot map cannot tell those apart, and they are 3.0 points of shooting percentage different.

15.9%fresh possession, from the slotno attempt of their own in 25 seconds
12.9%sustained pressure, from the same slottheir own last attempt 5 to 25 seconds ago
29.3%and a scramble, from the same slot22,985 attempts

Attempts inside 40ft and within 12ft of the middle lane
How the puck arrived	Shots	Goals	Conversion	Expected	Average distance
Fresh — a rush	131,482	20,888	15.9%	15.7%	18.0 ft
Sustained — a cycle	35,485	4,586	12.9%	13.0%	17.3 ft
A scramble	22,985	6,743	29.3%	29.1%	13.7 ft

What to do about it. Judge a club’s offence by where its chances come from as well as where they are taken from. A club that manufactures slot shots off the rush is generating something different from one grinding the same shots out of a cycle, and the difference is about a fifth of the value of each chance.

A shots-only feed cannot see a rush, so read these as shapes

Nothing here observes a zone entry, because the AHL publishes none. What is observed is the clock: how long it had been since that club last put a puck on net. A first attempt after twenty-five quiet seconds, taken from the slot, is shaped like a rush and converts like one; an attempt five to twenty-five seconds after their own last one is shaped like a cycle. Some rushes will be misfiled as cycles and the reverse, which if anything makes the gap here an understatement.

The three buckets partition every attempt exactly once, and the expected column tracks each of them — the model reads the same gap. The point is not that the model is surprised; it is that a reader looking at a shot map is, because the map draws all three as the same dot in the same place.

How the puck arrived beats where it was shot from

Three shots from the same patch of slot, three different games.

A goaltender's season barely predicts his next one

A forward’s scoring carries into the following season at r = 0.67. A goaltender’s save percentage carries at r = 0.29 — almost exactly the same as a forward’s shooting percentage (0.26), the one number nobody projects forward. Percentages do not repeat for anyone; a skater simply has volume as well, and a goaltender has nothing else at all.

0.67year-over-year r, forward points per game1,107 season pairs, 40+ games each
0.29year-over-year r, goaltender save percentage230 season pairs, 400+ shots each
0.26and a forward’s shooting percentagethe number everyone already knows not to trust

Correlation between one season and the next, same player
What repeats	Who	r	Season pairs
Expected goals per game	Forwards	0.690	1,107
Points per game	Forwards	0.666	1,107
Shooting percentage	Forwards	0.257	1,107
Goals saved above expected, per shot	Goaltenders	0.281	230
Save percentage	Goaltenders	0.285	230
Rebound leak rate	Goaltenders	0.144	230
High-danger leak rate	Goaltenders	0.050	230

What to do about it. Never sign, start or drop a goaltender on one season of anything, and be most suspicious of the stats that sound most diagnostic — a high-danger leak rate off a hundred charted saves is noise wearing a lab coat. The talent board is built for exactly this: every rating shrunk toward average by what the sample can support, with the interval that is left drawn on.

Thresholds, and why forwards only

Consecutive seasons for the same player, by start date — the season identifiers in this database are not sequential. A skater needs 40 games in both seasons and a goaltender 400 shots in both, which are different bars because they measure the same thing: enough of a workload that the number is about the player.

Defencemen are left out of the skater rows on purpose. Pooled with forwards, shooting percentage appears to repeat at 0.59 — but that is almost entirely the gap between the two groups, since defencemen shoot from the point every year and forwards do not. Within position it is 0.23 to 0.32 everywhere, and the pooled figure is measuring positions rather than players. The goaltender rows have no such problem; there is only one kind of goaltender.

The honest limit. A correlation is not a ceiling on what is knowable — it is what this metric, over this sample, carried forward. Shrunk career figures do better than any single season here, which is the whole argument for reading them that way, and none of this says a goaltender cannot be good. It says a season is a thin piece of evidence that he is.

A goaltender's season barely predicts his next one

Volume repeats and percentages do not — and a goaltender has nothing but a percentage.