The Right to Stop a Balanced Deck Is Worth $2.62

A shuffled deck of fifty-two is turned face up one card at a time. Every red card pays you a dollar, every black card costs you one, and you may walk away after any card. What is that right to walk away worth?

The deck holds twenty-six of each colour, so turning all of it lands on exactly zero. That cancellation is real, and it is the reason the answer is not zero. Zero is what you get for free; the option to stop can only add.

Zero is the floor, not the value

A strategy that turns every card scores $26 - 26 = 0$ with certainty. Any player who does at least that well in every deal has a value of at least zero, and stopping early is available at every moment, so the value is at least zero and strictly more as soon as one stopping rule beats it in some deals and loses in none. Such a rule exists at four cards already, which is small enough to check by hand.

There is a ceiling too, and it is worth having in mind for scale. A player who could see the whole shuffle in advance would stop at the running maximum, and no realisation of that maximum exceeds twenty-six. The blind player with a stopping right gets 2.62, so the information is worth a great deal more than the option is.

The recursion, from the last card backwards

From $(r,b)$ the next card is red with probability $r/(r+b)$ and black otherwise, and the player compares stopping against continuing:

The zero inside the maximum is the stopping payoff, which is nothing further, not nothing overall: what you have already banked is yours. Fill the table from $(0,0)$ outward and every entry needs only entries with one fewer card, so the whole $27 \times 27$ grid is one sweep.

The smallest case where the option is visibly worth something is one red and one black. The continuation from $(1,1)$ is $\tfrac12(1 + E(0,1)) + \tfrac12(-1 + E(1,0))$, with $E(0,1) = 0$ because the only card left is black and you refuse it, and $E(1,0) = 1$ because the only card left is red and you take it. That gives $\tfrac12(1) + \tfrac12(0) = \tfrac12$. Half a dollar, out of a two-card deck that also cancels to zero.

The exact value of the full deck

Run equation (1) in exact rational arithmetic and the fifty-two-card game has a value that is not an estimate:

Rounded to the cent, $2.62. The fourteen-digit numerator is what backward induction over a deck actually produces, and the fact that it is a rational number at all is a reminder that nothing here was sampled. A Monte Carlo over two hundred thousand deals under this policy measured 2.6203, which is the sort of agreement you want from a simulation whose job is to check rather than to compute.

The small decks are the useful part of that picture, because they can be verified by hand in a minute and they show the value climbing slowly. Doubling the deck from four cards to eight adds only a third of a dollar. Whatever the growth rate is, it is a long way below linear, which is what you should expect from something that scales like the typical excursion of a balanced walk rather than like its length.

The optimal policy cannot finish in the red

Here is the sharper statement, and no closed form announces it. Under the optimal rule the realised payoff is never negative, in any deal, not merely on average. Two facts do it.

First, playing to exhaustion from $(r,b)$ yields exactly $r - b$, so $E(r,b) \ge r - b$. Second, $E(r,b) \ge 0$ straight from the maximum in (1). Now suppose the policy stops at $(r,b)$, which means $E(r,b) = 0$. Combining the two facts forces $r \le b$. And the score already banked at that moment is:

Every card turned so far was red or black, so the tally is the reds seen minus the blacks seen, which is what (3) computes. The player only ever stops with more blacks gone than reds, which is the same as being level or ahead. Every stopping state in the grid was then checked against $r \le b$ and all of them satisfy it, and a simulation under the exact policy recorded a worst outcome of 0 and no negative deals at all.

Rules that sound optimal and are not

The obvious heuristic is to stop the first time you are ahead. It is not optimal and the gap is large: over the same deals it earned 0.9630 against the optimum’s 2.62. Stopping at plus one throws away the fact that a deck rich in red cards is still rich in red cards, and the state $(r,b)$ is precisely the thing that tells you so.

The other tempting rule is to keep going while reds outnumber blacks in the remainder, which is the same as stopping everywhere on the shaded half of Fig. 2. That has the right shape and the wrong boundary. The table says the player stops in strictly fewer states than the half contains: with one red card left he continues until two blacks remain, with two reds left he needs four blacks, and with ten reds left he needs fourteen. A bare surplus of blacks is not enough to make him quit, and the surplus he demands grows by roughly one and a half cards for every red still buried. No short formula produces those numbers, which is the honest reason to run the table rather than reason about it.

What the model needs

Three things, each of which changes the answer if you remove it. The shuffle must be uniform, or the count of remaining cards no longer determines the next card’s colour. Every turned card must be visible, since the whole value comes from conditioning on what has gone. And there is no discounting or cost per card, so waiting is free. Charge even a penny a card and the stopping boundary moves toward the start of the deck immediately.

Sources and further reading

• The general framework this problem is an instance of — Optimal stopping
• Filling the table from the last card backwards — Backward induction
• Why the recursion is cheap despite the size of the strategy space — Dynamic programming
• Sampling without replacement, which is the reason the state is two numbers — Hypergeometric distribution

The value in equation (2) is a proof by computation rather than a simulation: the recursion was run in exact fractions over the whole grid with no floating point anywhere. The never-negative property was proved as above and then checked at every stopping state, and two controls were run alongside, one playing every card and one stopping at the first profit.