Optimal Stopping Theory

An American option holder watches the underlying move every day and must decide, on each day, whether to exercise now or keep waiting. A trader holding a winning position faces the same shape of question: take the profit now, or hold for more, knowing the position could also reverse. Optimal stopping theory is the general framework for exactly this class of problem — deciding when to act on a random process, not what action to take.

An analogy: selling a house with random offers

Imagine you're selling a house and offers arrive one at a time, in random order and random size, and you must accept or reject each one on the spot — no going back to a rejected offer later. Accepting locks in that price; rejecting means waiting for the next, uncertain offer, possibly at some cost (time, carrying costs) for waiting. The optimal strategy isn't "take the first good offer" or "hold out for the best possible offer" — it's a threshold that itself depends on how many offers you still expect to see and how much waiting costs, and it can be worked out precisely.

The mathematics, one symbol at a time

Let $X_t$ be the random process being observed (an offer sequence, a stock price), and let $\tau$ be a stopping time — a moment to stop that depends only on information available up to that moment, not the unknown future (see Stopping Times). The optimal stopping problem is to choose $\tau$ to maximize expected reward:

$$V(x) = \sup_{\tau} \; \mathbb{E}_x\big[ g(X_\tau) \big],$$

where $g$ is the payoff for stopping at a given state, and the supremum is over all valid stopping times. In words: among every sensible "when to stop" rule you could follow (using only information you actually have at the time), find the one that gives the highest expected payoff, and $V(x)$ is the best achievable value starting from state $x$. The classical solution method splits the state space into a continuation region (keep waiting) and a stopping region (act now), separated by an optimal boundary; you stop the first time the process crosses into the stopping region. For American options this boundary is the early-exercise frontier; for the house-selling problem it's the acceptance threshold.

Worked example 1: a simplified selling problem

Suppose only two offers will ever arrive, each independently $80k with probability 0.5 or $120k with probability 0.5, and you must decide on the first before seeing the second. If you reject the first offer, you get whatever the second offer is, with expected value 0.5(80) + 0.5(120) = \100k. So the optimal rule is: accept the first offer only if it's at least \100k (i.e., accept $120k, reject $80k and wait for the guaranteed-in-distribution $100k average). Under this rule, expected payoff is 0.5(120) + 0.5(100) = \110k$— better than always taking the first offer ($$100k$expected) or always waiting for the second ($$100k$ expected).

Worked example 2: American put early exercise, one step

A one-period binomial model: stock at $100 can go to $120 (up, probability 0.5) or $85 (down, probability 0.5) at expiry, risk-free rate 0%. An American put struck at $100 pays $\max(100-S,0)$. If exercised immediately, payoff is $0 (at the money). If held to the single future date, expected payoff is 0.5(0) + 0.5(15) = \7.50$. Since$7.50 > 0$, the optimal action is to not exercise now — continuation beats immediate stopping here, exactly the comparison optimal stopping formalizes at every node of an American-style pricing tree.

What this means in practice

Optimal stopping is the theoretical foundation of American option pricing (the early-exercise boundary is exactly this continuation/stopping split), and shows up directly in interview brainteasers like the secretary problem (see The Secretary Problem). In trading, informal "take-profit" and "stop-loss" rules are crude, hand-tuned approximations to an underlying optimal stopping problem that a more careful model could, in principle, solve exactly given a payoff function and process dynamics.