Every online tournament lobby offers the same quiet bet. You can sit down at level 1 with a full stack and a soft table, or you can turn up two hours later, pay the same money, and receive a stack that is already short. Most serious players have heard that the second option is free equity. The Independent Chip Model agrees with them: buy in once half the field has busted and your stack is worth more than you paid for it before you look at a card.
We wanted to know whether the table actually pays that out. So we built a multi-table tournament simulator, calibrated it against published field data, and ran 586,000 tournaments through it. Inside every one of them sat a cohort of identical players who differed in exactly one respect: the level at which they registered.
Four things came back, and only the second is the one everybody talks about.
- The structural transfer is real, and smaller than advertised. In a field where every player is equally skilled, so that nothing but the structure can move the number, entering at the close of registration was worth between +3.9 and +11.4 ROI points across every setting of the one parameter we could not look up anywhere. Always positive, never as large as ICM claims.
- ICM gets the size right and the question wrong. At the close of late registration ICM values a fresh starting stack at +7.0% over the buy-in, and in a controlled equal-skill field the table paid +6.5%, which is the same number within measurement error. But ICM produces that one figure no matter who is holding the stack, while the delivered value moves across the whole sweep. It is not inaccurate here so much as incomplete.
- The weaker the player, the more late registration is worth. Registering as late as possible instead of at level 1 moved a recreational player's ROI by +9.9 points and a solid regular's by +2.4. It moved a professional's by -17.3. Most of what late registration does for a losing player is shorten the game.
- For a winning player the real prize is the clock, not the equity. Our professional earned $35.09 per hour entering at level 1 and $61.19 per hour entering at the close, because the second tournament takes 38 minutes instead of 99. Registering late costs a strong player ROI and pays them back in hours, at 1.7 times the rate.
This is the long version, and it is long: the model, the controls, the confidence intervals and the two bugs we had to fix before the numbers meant anything. If you want the practical answer in ten minutes instead, read should you late register and come back here when you want to argue with the method.
The rest of this article shows the working: what ICM claims, why it claims it, where the claim breaks, and what 586,000 simulated tournaments say instead. Every chart is generated from the run data, one of our own bugs is described because it changed the answer, and the assumption we were least sure of is swept on a chart rather than buried in a footnote.
What Does ICM Say Late Registration Is Worth?
ICM says a fresh starting stack is worth more than one buy-in as soon as anyone has busted, because the eliminated players forfeited prize equity and that equity was redistributed to everyone still alive. Buy in afterwards and you pay the old price for a share of a pot that has already been concentrated.
The cleanest way to see it is on a single table. Nine players sit down in a tournament paying 50 / 30 / 20. Four of them bust, their chips absorbed by four opponents who doubled up. Now you buy in as the tenth entrant with a starting stack, and your buy-in is added to the prize pool. Six players, one prize pool, and you own the shortest stack at the table.
Run the exact Malmuth-Harville recursion on that position and your seat is worth +8.7% more than you paid. You have not played a hand. If you want the mechanics of that calculation, our ICM explainer works through a three-handed example by hand.
| Payout structure | 0 busted | 1 busted | 2 busted | 3 busted | 4 busted |
|---|---|---|---|---|---|
| Standard (50 / 30 / 20) | +0.0% | +1.9% | +4.0% | +6.3% | +8.7% |
| Flat (5 places paid) | +0.0% | +4.2% | +9.3% | +16.1% | +26.3% |
| Steep (65 / 25 / 10) | +0.0% | +1.3% | +2.6% | +4.0% | +5.5% |
| Satellite (4 equal seats) | -0.0% | +4.2% | +9.1% | +15.2% | +22.6% |
| Winner-take-all | +0.0% | +0.0% | +0.0% | +0.0% | +0.0% |
Notice the bottom row before anything else.
Why the ICM Boost Is Not Free Money
In a winner-take-all tournament the late-registration boost is exactly zero, at every stage, forever. That single fact tells you what the boost actually is. It is not value created by waiting. It is value transferred out of the payout ladder, and if the ladder has only one rung there is nothing to transfer.
The logic is worth being precise about, because it is the foundation for everything below. In a winner-take-all event your equity is exactly your share of the chips. A starting stack is a starting stack whether it arrives at level 1 or level 12, and adding your buy-in to the prize pool adds exactly the same fraction that your chips add to the chip pool. Nothing moves.
The moment the structure pays more than one place, chips stop converting to money at a constant rate. Survival acquires cash value of its own, short stacks become worth more per chip than the chip leader's chips, and a player who buys a below-average stack into a shrunken field is buying at a discount to that ladder. The flatter the ladder, the bigger the discount: +26.3% for a structure paying five places, +22.6% for a satellite paying four identical seats, against +5.5% for a steep one.
Two consequences follow immediately, and both are usually left out when this number gets quoted.
First, somebody pays for it. The equity you collect by arriving late was surrendered by the players who busted before you arrived, and it is diluted for everyone who was already seated. Late registration is not a free lunch discovered in the rules. It is a transfer, and you are on the receiving end only because other people are on the paying end.
Second, the usual number is measured wrong. Quote the boost halfway through late registration and you are counting a prize pool that includes hundreds of entrants whose chips are not on the table yet. Their money is in the pool inflating your equity; their stacks are not in the chip count diluting it. Put those future entrants back in at their starting stacks and the picture changes a great deal.
The blue line is what you get if you run ICM on the players you can see. At level 1 it claims a starting stack is worth +151% more than the buy-in, which is obvious nonsense: it is dividing a 500-entry prize pool among the 40% of the field who have sat down so far. Halfway through the window it still claims +11.8%. Put the missing entrants back in at their starting stacks and the real figure at that moment is +1.5%.
The two lines meet at the close of registration, when there is genuinely nobody left to come, at +7.0%. That convergence is the sanity check, and the gold line is the honest number. It is the one we use for the rest of this article.
What the Published Numbers Say
The literature is small, remarkably consistent about the size of the ICM effect, and almost silent about whether that effect survives contact with a real tournament.
GTO Wizard's analysis of real tournament samples puts the boost at 9 to 12% in regular non-bounty events, measured where late registration closes with 37 to 48% of the field still alive: a $50 MTT at 37.5% remaining came out at +9.68%, a $15 MTT at 35% remaining at +11.89%. The same sample contains the two extremes that make the point: a satellite closing with only 24% of the field left was worth +57%, and a progressive knockout was worth -14%.
Their per-big-blind figure is the one worth remembering. In that $50 event a starting stack was worth 50 cents per big blind at the start of the tournament and $6.84 per big blind at the close of registration. The chips are identical. What changed is how large a share of how small a field they buy.
Dara O'Kearney, who wrote the book on ICM in tournaments, arrives in the same place from a different direction: about 10% for a ten-player sit-and-go, around 22% for entering a 33-player event on day two, and close to double the buy-in for registering on the literal money bubble. His conclusion is that late registration is clearly profitable, immediately followed by the warning that a skilled player usually earns more by playing from the first hand against weaker opposition.
Two studies go past the arithmetic. A Monte Carlo of one million simplified tournaments, giving professionals a 52-48 edge in every confrontation, produced ROIs of +15.2% for early-registering pros, +27.9% for late-registering pros and -18.4% for recreational players. Its own author flagged the fragility: raise the professional edge to 60-40 and the two registration strategies come out level, because that model gives skill nowhere to compound.
Separately, a study of real results across 1,004 players and 8,834 entries in Triton events, using each player as their own control to cancel skill differences, found ROI improvements of +13.2 points for entries made with 60 to 70% of the field remaining and +10.5 points below 60%. Its cleanest signal was not ROI but cash rate: players who started from level 1 cashed about 15% of the time, late registrants upwards of 19%. Hold on to that one, because our simulation reproduces it independently.
Against all of that, the practical advice from coaches runs the other way. Upswing's published guidance is to enter freezeouts with 40bb or more and knockouts with 50bb or more, and to enter turbos only at the start. Tournament regulars interviewed on the subject describe jumping in during the final minutes as straightforwardly losing, with one agreed exception: an overlay, where a guarantee has not been met and the prize pool contains money nobody paid for.
So the theory says late, the practitioners say early, and every source is arguing about the size of a number produced by a model that does not contain a big blind.
What ICM Cannot See
ICM converts stacks to money using two inputs, the stacks and the payouts, and nothing else. It has no concept of position, no concept of skill, and, decisively for this question, no concept of the blinds. It prices a 13bb stack and a 130bb stack identically if they hold the same number of chips.
That would be a harmless simplification if the error were random. It is not. It has a direction, it has been measured, and the direction is precisely the one that matters here.
- An empirical validation of ICM presented at the 2025 IEEE Conference on Games, fitted to results from over ten thousand real tournaments, reports "evidence of the ICM underestimating the performances of players with larger stacks while overestimating those who are short-stacked." That is the entire mechanism of this article, found independently in field data.
- A Statistical Science paper by Diaconis and Ethier compares ICM against the exact gambler's-ruin solution and finds it formally inadequate, with the disagreement growing without bound as a stack gets shorter. On the 2019 WSOP Main Event's final three, an ICM chop hands the short stack roughly $54,000 more than the exact model says it is worth.
- The Dependent Chip Model paper puts a number on the same bias in deal-making terms: in a three-handed example it prices the shortest stack 33.7% below what ICM would hand them in a chop.
- A thesis benchmarking equity models against a computed Nash equilibrium finds Malmuth-Harville ICM carries a mean absolute percentage deviation of 7.2%, the best of the plain equity models, and that Future Game Simulation roughly halves it, to 5.3% looking one hand ahead and 3.47% looking three. The thing FGS adds is exactly the thing ICM is missing: the cost of the blinds you are about to post.
Read those together and the late registrant's position is uncomfortable. Every published estimate of the late-registration boost is computed with a model that is known, from real data, to overvalue exactly the kind of stack a late registrant buys. The bias does not merely add noise to the answer. It inflates it.
Here is the same point in the numbers from our own simulated fields. As late registration runs, two things happen at once, and they pull in opposite directions.
The blue curve is what ICM sees and likes. It peaks while registration is still open, then turns over as eliminations outpace arrivals, and by the close only 47% of the eventual field is left to outlast.
What ICM prices badly is everything that curve does not show. The average surviving stack has climbed to 2.12 starting stacks, with a chip leader on 8.3, so the stack you just bought is only 47% of the table average. It is 13 big blinds rather than 100. And the players who busted to create that shrunken field were disproportionately the ones you wanted to play against.
It is worth ruling out the obvious suspect first. You might expect the late registrant to be punished by ICM pressure, the way a medium stack is punished on a bubble. They are not, and the numbers say so clearly.
The bubble factor of a freshly seated starting stack against a median covering stack rises only from 1.06 to 1.08 across the whole late registration window. At the close you need 51.8% equity to take a flip that chip maths prices at 50%. That is a real tax but a trivial one, and it agrees with GTO Wizard's own observation that bubble factors do not move much this early even though strategy does. The money bubble, hundreds of eliminations later, is where those numbers get violent, and our final table ICM guide covers what they do to ranges there.
So the late registrant's problem is not that ICM is squeezing them. It is that ICM is flattering them. And the reason sits in the same calculation.
This is the concavity of tournament equity, measured on real simulated fields at the close of late registration, and it explains both halves of the answer at once. A fresh starting stack is worth 1.07 buy-ins. Doubling it returns 2.06, which is a multiple of 1.93x, so at that depth chips convert to money almost linearly. Go up to twelve starting stacks and the same chips return 9.4 buy-ins instead of 11.7, a multiple of 8.75x.
Read that curve from the late registrant's seat. You are buying in near the bottom of it, in the stretch where an extra chip is worth close to face value, which is precisely why the ICM boost you collect is modest rather than enormous. The steep discount lands on the chip leader, not on you. What you are actually buying is a cheap ticket into a shrunken field, and the price of the ticket is that the blinds are already large enough to take it away from you.
It is worth turning that around, because the mirror image is the thing most players actually want to know: what is a big stack worth? On the pure ICM reading, less than you would hope. The chip leader at the close of registration is holding 8.3 starting stacks, and the curve says the marginal chip up there is worth a fraction of the marginal chip at the bottom. That is the honest arithmetic of the payout ladder, and it is why "I have to gamble to build a stack" is a much weaker argument than it feels like at the table.
But this is also exactly where ICM is known to be wrong in the other direction. The same empirical study that found ICM overestimating short stacks found it underestimating large ones, and the reason is not mysterious. A covering stack applies pressure that ICM has no way to represent, because ICM does not know that anybody has to make a decision. Opponents' calling ranges tighten against it, its steals get through more often, and none of that appears in a model whose only inputs are stacks and payouts.
So the true value of a big stack sits somewhere above the gold line and below the blue one, and the gap between them is made of playing skill. That is the same sentence as the one about short stacks, pointed the other way, and it is why this article ends up being about how you play rather than about when you sit down.
How We Simulated It
The question we needed to answer is not one ICM can answer, because ICM is one of the things being tested. So we built a tournament instead.
It is worth naming what that is. Future Game Simulation improves on ICM by playing the next one to three hands forward before pricing the position, and the thesis quoted above shows that doing so halves ICM's error. What follows is the same idea taken to its limit: rather than looking three hands ahead, we play every remaining hand of the tournament, hundreds of times per entrant, and read the answer off the payouts. That is expensive, which is why nobody uses it at the table, and it is exactly right for settling an argument.
The simulator is a chip-transfer model rather than a hand-by-hand poker engine, which is the right level of abstraction here: simulating individual streets would add enormous cost and no accuracy to a question about entry timing. What it does model is the four things that actually drive the answer.
- Chips are conserved. Every chip won came from another player. Blinds and antes form a dead pot that somebody collects.
- The blind clock escalates. Level 1 is 100bb deep and the close of registration is 13bb deep, with a big blind ante throughout. This is the mechanism ICM is blind to, so it had to be explicit.
- Skill converts to chips faster when stacks are deeper, saturating around 40bb. A good player at 100bb realises materially more edge per hand than the same player at 10bb, which is the standard argument for registering early and had to be in the model for the test to be fair.
- Elimination is terminal and the ladder is non-linear. A 500-runner field, 75 places paid (15% of the field), a harmonic payout ladder that reproduces real online structures closely: the winner takes 20.4% of the pool and the min-cash pays 1.4 buy-ins.
Every tick, all surviving players are randomly re-seated into nine-handed tables and one hand is dealt everywhere. Blinds and antes are posted into a dead pot, which is awarded by a weighting over skill and stack depth. Separately, with some probability, two players contest a pot whose size is drawn from a heavy-tailed distribution capped by the effective stack, so most pots are small and a few are stack-offs. Players below 12bb are pulled into pots more often, as short stacks are in real games.
Keeping those two things separate matters more than it sounds. In an earlier version of this engine the two players in a contested pot also split the dead money between them, which quietly made simply being in a pot worth four times a fair share. Because short stacks are in more pots, that handed them free chips, and free chips to short stacks is exactly the bug that manufactures a late-registration edge out of nothing. The engine now ships with an invariant test that fails if a player's expected chip change depends on their stack size when every player is identical.
The free parameters were not chosen to make a point. They were fitted to published targets before any entry-timing question was asked, and then frozen.
| What we tied the model to | Target | What the model produced |
|---|---|---|
| Share of the field still alive when late registration closes | 42 to 48% (real samples) | 47% |
| Average surviving stack, in starting stacks | about 2x | 2.13x |
| Chip leader then, in starting stacks | 8 to 12x | 8.3x |
| Recreational ROI before fees | about -20% | -25.9% |
| Regular ROI before fees | about +4% | +8.6% |
| Professional ROI before fees | about +49% | +54.7% |
| Whole field, before fees (must be exactly zero) | 0.00% | -0.00% |
Registration is staggered: recreational players front-load their entries and professionals spread out, which is what the field data implies and which, as it turns out, matters a great deal. Field composition is held on target so the test cohort never quietly toughens or softens the field it is being measured in.
The experiment itself is a within-tournament design. Every tournament seeds five identical copies of the same player at each of the 12 possible entry levels, sixty test players in a field of 500. All sixty share one tournament, one field and one set of opponents, so the only thing that differs between them is when they sat down. Any run of luck in how that particular field turned out hits all twelve cohorts at once.
On error bars: tournament results are violently skewed, so we quote them for the quantity that matters. Each headline figure rests on at least 275,000 simulated entries, and the difference between two entry levels, which is what every claim below is about, carries a standard error of roughly 2.0 points. We computed that error across tournaments rather than across entries, because the five clones seeded at one level share a field and are not independent observations. Doing it the easy way would have made our confidence intervals look about 60% tighter than they are.
Result 1: The Structural Transfer, and How Much to Trust It
To test ICM's promise cleanly we ran a version of the tournament in which every player is equally skilled. No professionals, no recreational players, no soft early field, nothing that can move the result except the structure itself. In that world the whole field averages exactly the buy-in before fees, so any entry level returning more than that is measuring the structural value of arriving at that moment and nothing else. We assert it as a test: the engine fails its own suite if that average drifts.
Before trusting any of it, though, we had to deal with a parameter we could not look up anywhere. The engine has to decide how often a shoving short stack takes down the blinds and antes. A player with no shove leverage whatsoever wins the dead money exactly as often as any other seat, which is 11% of the time at a nine-handed table. A player whose shoves are feared wins it far more often. Nothing published pins the real figure down, and it is precisely the quantity a late registrant lives or dies by, so we swept it end to end.
The result is the reassuring kind. Across the entire plausible range, from a short stack with no leverage at all to one that takes the pot 36% of the time, the structural value of registering at the close never turns negative: it runs from +3.9 to +11.4 ROI points against registering at level 1. The size of the prize moves; its existence does not. More shove leverage helps, by roughly 2.1 ROI points for every ten percentage points of shove success, but it never changes the sign and never comes close to the size of the effects in the next three sections.
Individual points carry a standard error of roughly 2.9 points, which is why the chart shows a fitted line through them rather than pretending each dot is exact. That noise is also why we are not going to tell you the transfer is worth some precise figure: what the data supports is a range, and the range is comfortably positive.
Running the same sweep on a professional in a normal mixed field says something the equal-skill version cannot. Their penalty for entering late shrinks from about -14.3% at the floor to about -3.6% at the top. A strong player who is genuinely good with ten big blinds gives up far less by turning up late than one who is not, which is the practical version of this whole article and the reason the last section is about push-fold rather than about clocks.
For every result below we pin the dial at the far left of this chart: a short stack wins the blinds 11% of the time, its bare seat share, with no credit whatsoever for the fact that people do not enjoy calling all-ins. That is certainly too harsh. We use it anyway, because it is the one setting that cannot be accused of being chosen to produce a result, and because the sweep has just shown that the choice does not swing the answer. Everything that follows is a floor.
Result 2: ICM Called the Size, and Cannot Call the Player
Now hold ICM's prediction against the outcome, in the same units, on the same fields. The interesting part is not the size of the gap. It is that ICM does not have a gap that can be quoted.
The gold line is ICM's answer. It rises smoothly to +7.0% at the close of registration and it is the same line no matter who you are, because the model's only inputs are stacks and payouts. The two other lines are what the table actually paid in an equal-skill field, at the bottom and the top of the shove-leverage sweep. Reality is not one line near ICM's. It is a band that ICM sits somewhere inside.
| Enter at | Stack | ICM predicts | No shove leverage | At the neutral setting | Strong shove leverage |
|---|---|---|---|---|---|
| Level 1 | 100bb | -0.0% | +0.0% | +0.0% | +0.0% |
| Level 2 | 83bb | +0.1% | +3.9% | +1.5% | +0.2% |
| Level 3 | 69bb | +0.4% | +3.9% | +0.8% | -1.6% |
| Level 4 | 58bb | +0.7% | +3.3% | +3.2% | -2.5% |
| Level 5 | 48bb | +1.1% | +1.9% | +2.7% | -3.8% |
| Level 6 | 40bb | +1.5% | +7.2% | +3.2% | +3.6% |
| Level 7 | 33bb | +2.1% | +2.1% | +2.4% | -2.7% |
| Level 8 | 28bb | +2.7% | +2.0% | +2.4% | +1.4% |
| Level 9 | 23bb | +3.5% | -0.0% | +3.7% | +4.9% |
| Level 10 | 19bb | +4.4% | +7.5% | +4.7% | +3.3% |
| Level 11 | 16bb | +5.5% | +8.5% | +5.8% | +7.4% |
| Level 12 | 13bb | +7.0% | +4.2% | +6.5% | +12.2% |
Here is the number we expected to be able to attack, and could not. At the floor setting, entering at the close was worth +6.5% more than entering at level 1, with a standard error of 1.4 points, against ICM's prediction of +7.0%. The gap is +0.5 points, which is 0.4 standard errors. ICM called it. We went looking for a systematic overstatement, on good theoretical grounds and with an empirical paper pointing the same way, and in this particular decision it is not there.
Two things explain why the bias does not bite here, and both are already on charts above. The late registrant's stack sits in the near-linear stretch of the equity curve, where an extra chip is worth close to face value, so the concavity that trips ICM up is barely engaged. And their bubble factor is 1.08, meaning ICM pressure at this point in a 500-runner field is almost nothing. ICM is least reliable for short stacks near a pay jump. A late registrant is a short stack a very long way from one.
What ICM still cannot do is tell those two players apart. At the optimistic end of the sweep the same decision was worth +12.2%, comfortably more than ICM promised, and ICM's answer did not move, because stacks and payouts are its only inputs. It is not wrong by a percentage here. It is answering a question with one number that has more than one answer.
That is worth saying plainly, because it cuts against the direction we set out expecting. The 2025 IEEE Conference on Games paper does find ICM overestimating short stacks on real tournament data, and we believe it. What our numbers add is where that bias does and does not matter: it is a bias about stacks near elimination and near a pay jump, and the late-registration decision sits in neither place. Quoting the ICM boost for max late registration is fine. Quoting it as though it applied equally to every player is not.
Does the transfer pay the rake?
This is the version of the question that actually decides whether a marginal player is a winner, and it is worth doing explicitly because every ROI figure in this article is quoted after fees.
Our tournament is $100 + $8, so the fee is 8% of the money that reaches the prize pool. Against that, the structural transfer from registering at the close measured +6.5% at the floor setting, rising to +12.2% at the top of the sweep. So even on the most conservative reading late registration claws back most of the tournament fee, and on a generous one it covers the fee outright with change to spare. GTO Wizard reach the same conclusion from their real-tournament sample, where a 9 to 12% boost "either negated or beat the rake" in every non-bounty event they measured.
That framing is the one to carry to the table. Late registration is not a strategy that makes you a winning player. At best it is a rebate on the entry fee, and how much of the rebate you collect depends on how well you play the stack it hands you.
Result 3: Who Actually Gains From Late Registration
Now put the skill back in and ask the question players actually care about. Same tournament, same field, three player types, each measured at all 12 entry points.
The three curves do not have the same shape, and that is the finding.
| Enter at | Stack | Field alive | Recreational | Regular | Professional |
|---|---|---|---|---|---|
| Level 1 | 100bb | 40% | -31.1% | +0.2% | +53.4% |
| Level 2 | 83bb | 62% | -31.8% | +1.5% | +44.6% |
| Level 3 | 69bb | 72% | -32.2% | -0.1% | +46.0% |
| Level 4 | 58bb | 77% | -29.6% | +0.5% | +45.5% |
| Level 5 | 48bb | 78% | -29.3% | +1.4% | +43.9% |
| Level 6 | 40bb | 76% | -31.4% | -0.9% | +41.0% |
| Level 7 | 33bb | 72% | -27.4% | -1.6% | +40.5% |
| Level 8 | 28bb | 68% | -29.5% | -0.2% | +40.4% |
| Level 9 | 23bb | 63% | -27.2% | -0.8% | +36.8% |
| Level 10 | 19bb | 58% | -25.8% | -0.6% | +37.5% |
| Level 11 | 16bb | 52% | -23.6% | +1.4% | +33.7% |
| Level 12 | 13bb | 47% | -21.2% | +2.6% | +36.1% |
For the recreational player, registering late is close to the most profitable decision available anywhere in this article. ROI after fees improves from -31.1% to -21.2%, a gain of 9.9 points, and it improves at almost every step along the way. Two forces push the same direction. They collect the ICM transfer like everybody else, and they spend 34 minutes at the table instead of 85, which is an hour of their own mistakes that never get made. A losing player registering late is buying less exposure to themselves.
The regular gains too, though more modestly and with less confidence: +0.2% at level 1 against +2.6% at the close, a swing of +2.4 points with a standard error of 1.5, so call it positive but do not bet the house on the exact size. Their cash rate moves more convincingly, from 15.8% to 17.6%. For a player whose edge is thin enough that the fee decides whether they are a winner, a couple of points is not nothing.
For the professional, the sign flips. Entering at level 1 returned +53.4%. Entering at the close returned +36.1%, a change of -17.3 points, and that is not noise. The structural transfer is still there and still positive; it is simply much smaller than what a strong player gives up to collect it.
Two things are being surrendered, and both are measurable in the run. The first is the soft field. The average skill of the players seated at level 1 is -0.506 on our scale against -0.211 at the close, because recreational players register early and then bust, so the strongest fields of the tournament are the ones you buy into last. The second is stack depth: 100 big blinds is where a skill edge compounds, 13 big blinds is where it is compressed into a shove-or-fold decision that a competent amateur can also get right.
So the headline reverses the usual framing. Late registration is not an edge that good players exploit and weak players miss. It behaves much more like a subsidy, and it is worth most to the people who are worst at poker, because most of what it does is shorten the game.
One thing moves the same way for everybody, and it is the cleanest signal in the study because it is far less noisy than ROI. Cash rate rises with a later entry. For the recreational player it went from 12.6% to 15.9%, and for the regular from 15.8% to 17.6%, because the field you have to outlast is smaller by the time you sit down. That is exactly what the real Triton data shows: players starting from level 1 cashed about 15% of the time against upwards of 19% for late registrants.
Win rate does not follow it, and for a strong player it moves the other way. Our professional won the tournament 0.45% of the time entering at level 1 and 0.38% entering at the close, and their cash rate stayed flat at 19.7% against 19.6%. Late registration buys min-cashes for the players who were not going to reach the top of the ladder anyway. It does not buy anybody a better chance of winning the thing, and the top of the ladder is where the money is.
Result 4: The Hourly Rate Changes the Answer
Everything above measures return per buy-in. Almost nobody's actual constraint is buy-ins. It is hours.
Our professional made $35.09 per hour of tournament time entering at level 1, and $61.19 per hour entering at the close, a factor of 1.7. That is the same player who just lost 17.3 points of ROI by making that choice. Both facts are true at once, and they are not in tension: the late tournament lasts 38 minutes rather than 99, so a smaller edge is being realised 1.57 times an hour instead of 0.61.
Which of those two numbers you should care about depends entirely on what is scarce for you. If you can reliably find more tournaments at your stake to fill the time you free up, the hourly figure is the one that pays your rent. If your volume is capped by the schedule rather than by your clock, it is a mirage, and the ROI column is the only one that matters.
| Enter at | Stack | Time at the table | Tournaments per hour | Professional ROI | Professional $/hour |
|---|---|---|---|---|---|
| Level 1 | 100bb | 99 min | 0.61 | +53.4% | $35.09 |
| Level 2 | 83bb | 92 min | 0.65 | +44.6% | $31.52 |
| Level 3 | 69bb | 86 min | 0.70 | +46.0% | $34.79 |
| Level 4 | 58bb | 80 min | 0.75 | +45.5% | $37.04 |
| Level 5 | 48bb | 73 min | 0.82 | +43.9% | $38.70 |
| Level 6 | 40bb | 68 min | 0.88 | +41.0% | $39.07 |
| Level 7 | 33bb | 63 min | 0.95 | +40.5% | $41.73 |
| Level 8 | 28bb | 57 min | 1.04 | +40.4% | $45.61 |
| Level 9 | 23bb | 52 min | 1.15 | +36.8% | $45.53 |
| Level 10 | 19bb | 48 min | 1.26 | +37.5% | $50.98 |
| Level 11 | 16bb | 43 min | 1.40 | +33.7% | $50.80 |
| Level 12 | 13bb | 38 min | 1.57 | +36.1% | $61.19 |
This is the piece that decides the argument for anyone playing volume, and it survives the objection you are probably forming. Most online grinders multi-table, which changes the arithmetic, but it changes it in the same direction: a shorter tournament frees a seat sooner, and the levels being skipped are precisely the ones where a big blind is worth the least real money.
The same chart carries a warning for the recreational player, and it is the most uncomfortable result in the study. Their ROI improved by 9.9 points, which is a real saving per buy-in. Their rate of loss per hour at the table got substantially worse: $-23.71 an hour registering at level 1 against $-40.66 an hour registering at the close. Registering late does not make a losing player lose more slowly. It makes them lose less per tournament by buying less tournament, and the tournament they do buy is the expensive part. If what you are trying to protect is your bankroll, take the deal. If what you were buying was an evening of poker, you have just paid a similar rate for a shorter evening.
One popular objection does not survive the data. Late registration is widely believed to be a high-variance strategy, on the grounds that you are buying a short stack. In our runs the standard deviation of a single entry's result was 7.6 buy-ins registering at level 1 and 7.0 buy-ins registering at the close, a change of -8%. An independent analysis of real tournament data reached the same conclusion, putting the change in standard deviation between a day 1 and a day 2 entrant at roughly plus or minus one percent. Cashing more often and winning no more often largely cancel out. If anything, playing more tournaments per hour pulls variance down over a season. Our bankroll guide covers the buy-in counts that follow.
Result 5: The Payout Structure Decides Almost Everything
Change nothing about the tournament except the shape of the payout ladder, and the value of late registration moves further than any other variable we tested.
| Payout structure | Places paid | Enter at level 1 | Enter at the close | Change |
|---|---|---|---|---|
| Satellite (50 equal seats) | 50 | +28.0% | +18.6% | -9.4 pts |
| Flat (25% paid) | 125 | +28.0% | +33.8% | +5.8 pts |
| Standard (15% paid) | 75 | +53.4% | +36.1% | -17.3 pts |
| Steep (10% paid) | 50 | +72.7% | +48.7% | -24.0 pts |
| Winner-take-all | 1 | +110.6% | +83.8% | -26.9 pts |
These are measured on the professional, so read them against that player's baseline rather than against zero: in the standard structure they lost -17.3% by registering at the close, for all the reasons the last two sections gave. The question here is how far each payout shape moves them from that baseline.
Two separate levers are doing the work, and players routinely collapse them into one. The first is how many places get paid. The second is how equal those payouts are. Our flat structure pays a quarter of the field on a gentle ladder and comes out at +5.8%; our satellite pays a tenth of the field in perfectly equal seats and comes out at -9.4%. The steep structure, paying only a tenth of the field on a very top-heavy ladder, comes out at -24.0%.
Winner-take-all is the control, and it comes back at -26.9%, the worst of the five. With one rung on the ladder there is no transfer available at any moment of the tournament, so what is left is purely the cost of a tougher field and a shorter stack. Treat that number loosely, mind: a structure that pays one player has ferocious variance and its error bar is about 14 points. The sign is not in doubt; the second decimal is meaningless.
The span from winner-take-all to the flat ladder is the clearest single measurement of the structural transfer in this whole study, because it is the same field, the same clock and the same player, with only the ladder changed.
The published figures behave the same way. GTO Wizard's real-tournament sample put a satellite at +57% against 9 to 12% for standard MTTs, and their satellite closed registration with only 24% of the field left, which is a third lever again: the later the window closes, the more of the field has already paid in and busted. Our satellite strategy guide covers why those structures bend every other decision too.
If you take one practical rule from this section, take this one: before deciding when to register, look at how many places get paid and how flat the ladder is. Those two numbers tell you more about the value of arriving late than your own win rate does.
Result 6: Bounty Tournaments Make It Worse Again
Progressive knockouts are the one format where the received wisdom is unambiguous: do not late register them. Our numbers agree, though by a smaller margin than the warnings in the literature imply.
In a PKO, half of every buy-in goes into a bounty pool that is paid out continuously as players are eliminated. Arrive late and you have missed a large share of it. The regular prize pool is half the size, so a min-cash that returns 1.4 buy-ins in a standard event returns well under one, and the ICM boost you collect is a boost on the smaller half of the prize money. Meanwhile the bounties you did not collect are gone for good.
Our simulation puts a professional at +28.5% entering at level 1 and +8.6% entering at the close, a change of -19.9%. Set that against the -17.3% the same player suffered in the standard event and the knockout format costs an extra 2.6 points on top of an already unattractive decision. The mechanism shows up directly in the payouts: bounties made up 43% of the level 1 entrant's total return and only 37% of the late entrant's. The money you came for has already been handed out.
The published analysis agrees on the mechanism: by the time late registration closes with 30 to 35% of the field remaining, the bounty pool is 40 to 45% depleted and the value of a starting stack has fallen about 14%. Mystery bounties are the exception that proves the rule, because their bounty phase typically starts on day 2, so there is nothing to miss.
Does the Answer Survive Different Assumptions?
Two of our modelling choices are assumptions rather than measurements, and both of them are load-bearing, so we re-ran the experiment with each one switched off.
| Model variant | Level 1 | At the close | Change |
|---|---|---|---|
| Baseline model | +53.4% | +36.1% | -17.3 pts |
| Everyone registers on the same schedule | +45.2% | +36.6% | -8.6 pts |
| Skill edge does not grow with stack depth | +69.7% | +61.0% | -8.7 pts |
| 1,000-entry field | +61.4% | +41.1% | -20.4 pts |
The first is who registers when. Our baseline has recreational players front-loading their entries and professionals spreading out, which is what makes the early field soft and is the main reason a strong player is rewarded for turning up at level 1. Give every player type the same registration schedule and that softness disappears by construction, which is the single largest lever on the professional's answer.
The second is whether a skill edge really grows with stack depth. This is the strongest argument for registering early and it is not calibrated to anything, so we removed it: same player, same edge per confrontation at 8bb as at 100bb.
Read together, the two runs decompose the professional's penalty almost exactly in half. Switch off the soft early field and 50% of it disappears. Switch off the depth advantage instead and 50% of it disappears. Neither assumption is carrying the result on its own, and neither is decorative: the reason a strong player should turn up at level 1 is both that the table is softer and that 100 big blinds is where their edge lives, in roughly equal measure.
That also tells you how to adjust for your own game. If your field has no recreational bias in registration times, halve the early-entry case. If your edge is mostly preflop and shove-shaped rather than postflop, halve it again, and late registration starts to look reasonable even for a winner.
What to Actually Do
The answer is not universal, which is why the question keeps getting argued about. It depends on four things, and the first one is the one the rest of this article was really about.
How good are you, relative to the field you are buying into? That turned out to be the variable that dominates everything else here. It decided whether registering late was worth +9.9 points or -17.3, a spread far larger than the structural transfer itself. Everything else on this list is a second-order adjustment to it.
- If you are a losing or breakeven player in the game you are playing: register late. This is the clearest result in the study. It reduced the recreational loss rate by 9.9 points, and it works by giving you the ICM transfer while cutting the number of hands in which you can hurt yourself.
- If you are a solid winner and your constraint is buy-ins: register at the start. On our numbers a professional gave up 17.3 points of ROI by turning up at the close instead, and the field is at its softest in the first two levels.
- If you are a solid winner and your constraint is hours: register late anyway, and do the arithmetic yourself. You will lose ROI per buy-in and gain a large multiple on your hourly rate. Whether that trade is good depends on whether you can actually fill the hours you free up with more tables at the same stake.
- Satellites and flat structures: late, emphatically. Steep structures and winner-take-all: the boost is small or absent, so decide on field quality alone.
- Progressive knockouts: early. The bounty pool is being spent while you are not there.
- Turbos and hyper-turbos: at the start. There is no late-registration window long enough to matter and no stack depth to work with once it closes.
- An overlay overrides all of the above. If a guarantee has not been met, the prize pool contains money nobody paid for, and taking a share of it beats every consideration in this article.
One more thing, and it is the part that turns this from trivia into a skill. If you late register, you are choosing to play a stack between 10 and 25 big blinds for the opening stretch of your tournament, and that is a different game with different correct answers. It is also the most solvable part of tournament poker, which makes it the cheapest edge in the study to actually acquire. Our short stack strategy guide covers the push-fold framework, and the ICM article explains why your calling ranges have to tighten when the covering stack is the one applying pressure.
Limitations
An honest data post says where its own numbers could be wrong, so here is the list.
This is a model, not a hand history database. The engine reproduces field decay, stack dispersion and ROI spread that match published figures, but it does not play poker. It cannot tell you that a particular 14bb shove is a fold.
The shove-leverage parameter is the weak point and we have not hidden it. It is not calibrated against anything, because nothing published pins it down, so we swept it rather than guessing. The good news is that the conclusion survives: the structural transfer stayed between +3.9 and +11.4 points across the whole range. We publish the figure at the floor of it, which makes every late-entry number here a lower bound rather than an estimate.
The skill-versus-depth relationship, the other load-bearing assumption, is also an assumption rather than a measurement, and it is the single most important one for the early-entry case. We tested it by switching it off rather than by defending it.
The field composition is a stylised three-type mix, and real fields are a continuum with correlations we have not modelled, including the strong likelihood that the players who register late are systematically different from those who do not, in ways beyond skill. Our design deliberately holds skill constant across entry cohorts to isolate the timing effect, which means it cannot capture that selection.
We wanted a frozen-blinds control to isolate the clock directly, and we do not have one. With the blinds never rising the field stops busting, the tournament does not finish inside any sensible compute budget, and finishing positions would have had to be assigned arbitrarily. We dropped it rather than report a broken control, which means the clock's contribution is argued here from the stack-depth numbers rather than demonstrated by removing it.
We model freezeouts. Re-entry changes the calculation, generally in favour of firing early enough to have bullets left. We model a single 500-runner field size with a single structure; the published data on field sizes and windows is consistent with our set-up, but a 40-minute-level live event and a hyper-turbo are different animals.
Finally, ICM itself is a model with known biases, and we use it in two roles here: as the thing being tested, and as the yardstick for the equity calculations. Where those two roles could contaminate each other, we used the simulation rather than ICM as the source of truth, which is why the controlled equal-skill experiment exists.
Common Questions
Is late registration actually +EV? The structural transfer is real and ICM sizes it correctly: in a field where every player is equally skilled, buying in at the close was worth +6.5% more than buying in at level 1, even after we stripped a short stack of any shove leverage, against an ICM prediction of +7.0%. But "is it +EV for me" is a different question, and the honest answer is that it depends on how well you play ten big blinds, how soft your field is early, and what your time is worth. This article exists because those three things move the answer more than the ICM transfer does.
How late is too late? It depends which player you are, which is the theme of the whole article. On our numbers a recreational player never reaches a point where arriving later is worse, so for them the answer is "there is no such thing as too late" in a non-bounty freezeout. A strong player is giving up ground from the first level onward. The common coaching advice, enter freezeouts with 40bb or more and knockouts with 50bb or more, is too conservative for a weak player in a standard MTT and about right for a strong one, which is roughly the opposite of how it is usually presented.
Does late registration raise variance? Essentially not at all, which surprised us. The standard deviation of a single entry's result changed by -8% between entering at level 1 and entering at the close, and an independent analysis of real tournament data puts the same figure at roughly plus or minus one percent. Cashing more often and winning no more often cancel each other out. What genuinely does reduce variance is playing more tournaments, which registering late lets you do.
Should I late register a satellite? Yes, more than anything else on this list. Flat payouts are where the ICM transfer is largest, and our simulation puts the gain at -9.4%.
Does this apply to live tournaments? The direction does; the magnitudes will differ. Live events have longer levels, so registering at the close leaves you deeper in big blinds than the 13bb our online structure implies, which strengthens the case for entering late because the stack you buy is more playable. The soft-field argument is usually stronger live too.
The Short Version
ICM is right that late registration is worth money, and on our numbers it sizes the prize accurately: it promised +7.0% and the table paid +6.5%. What it cannot tell you is the thing that actually decides your answer, because its only inputs are stacks and payouts and the question is really about who is holding them.
What the transfer does deliver goes disproportionately to players who were going to lose anyway, because the main thing a losing player buys by arriving late is fewer opportunities to play badly. Strong players registering late are not buying ROI, they are selling it: 17.3 points of it, in exchange for 1.7 times the hourly rate. Whether that is a good trade is a question about your calendar, not about poker.
Which points at the cheapest improvement available here, and it is not a change to your registration habits. Whatever time you sit down, at some point in every tournament you will be playing ten to twenty big blinds, and that is the most solvable stretch of tournament poker there is. Getting it right is worth more than getting the entry timing right, by our own numbers, by a factor of several. If you want to drill exactly those spots, that is what GTO Gecko was built for, and our training data post measures what it does to a real player's mistake rate.
Sources
- Kim (2025), Empirical Validation of the Independent Chip Model, IEEE Conference on Games
- Diaconis and Ethier (2022), Gambler's Ruin and the ICM, Statistical Science
- The Dependent Chip Model: a simple and more realistic alternative to the ICM
- Comparing equity models against a computed Nash equilibrium (TU Wien)
- GTO Wizard: Register Late, Win More
- GTO Wizard: How Significant Is ICM During Late Registration?
- GTO Wizard: The ICM Benefits of Late Registration
- GTO Wizard: How Payout Structures Impact ICM
- GTO Wizard: The Hidden Cost of Late Reg in PKOs
- GTO Wizard: When ICM Breaks Down
- PokerScout: Late-registering Triton players outperform their own earlier buy-ins
- PokerScout: Late-registering pros outperform equivalent pros in simulation
- PokerNews: Is Late Registration +EV? ICM Gives Us the Answer
- Upswing Poker: Late Registering Online Tournaments