7
$\begingroup$

Assume we are considering the discrete case. If $\{X_n\}_{n\in\mathbb{N}}$ is a martingale adapted to $F_n$ and $X_n\in L^1$, then for any bounded stopping time $\tau$, according to the optional stopping theorem (or optional sampling theorem), we will have that $E(X_\tau)=E(X_0)$.

I've seen some discussions on the condition such as the boundedness of the stopping time. However, now I am curious about that if we don't have the assumption that $X_n$ is a martingale does $E(X_\tau)=E(X_0)$ still holds for any bounded stopping time? In another word, does there exist any integrable random process $X_n$ adapted to $F_n$ which is not a martingale such that for any bounded stopping time $\tau$ it holds $E(X_\tau)=E(X_0)$?

Or we can rephrase this as "converse" of OST: if integrable random process $X_n$ adapted to $F_n$ and for any bounded stopping time $\tau$ it holds $E(X_\tau)=E(X_0)$, then $X_n$ is a martingale. Is this converse true or not?

$\endgroup$
7
$\begingroup$

Neat question! It is true, if I'm not mistaken.

Let $n$ be arbitrary and set $A = \{E[X_{n+1} \mid F_n] > X_n\} \in F_n$. Set $\tau = (n+1) 1_A + n 1_{A^c}$; you may verify that $\tau$ is a bounded stopping time. Then $X_\tau = X_{n+1} 1_A + X_n 1_{A^c}$.

Let $Y = E[X_\tau \mid F_n] = E[X_{n+1} \mid F_n] 1_{A} + X_n 1_{A^c}$ (by properties of conditional expectation). Note that $Y \ge X_n$, and $Y > X_n$ on $A$. But $E[Y] = E[X_\tau] = E[X_0] = E[X_n]$ since $n$ is also a bounded stopping time. So we conclude that $P(A) = 0$, which is to say $E[X_{n+1} \mid F_n] \le X_n$ almost surely. The opposite inequality can be shown in the same way, so we get $X_n = E[X_{n+1} \mid F_n]$ almost surely, and $n$ was arbitrary. Thus $X_n$ is a martingale.

Intuitively, if you think of $X_n$ as a stock price, then $A$ is the event that "information available by time $n$ suggests that, on average, the price is going to rise tomorrow". So $X_\tau$ corresponds to the strategy "if on day $n$, it looks like the price is going to rise, then hold until tomorrow; otherwise sell today". If $A$ had positive probability then this investment strategy would be profitable on average, which is supposed to be impossible if the stock price is a martingale.

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ Nice and clear answer. Thanks a lot. $\endgroup$ – Syoung Apr 8 '17 at 16:31
  • $\begingroup$ Brilliant answer, thanks for the informative text at the end ! $\endgroup$ – Gabriel Romon Oct 18 '17 at 11:30
4
$\begingroup$

Nate Eldredge's answer provides some nice intuition on why the result holds true. Here's a different approach.

The task at hand is to prove that $\forall n, E(X_{n+1}|F_n)=X_n$. Let $n$ be fixed. Going back to very definition of conditional expectation, it is sufficient to prove that for every $A\in F_n$, $E(1_A X_{n+1})=E(1_AX_n)$.

Define $\tau = n 1_A + (n+1)1_{A^c}$. $\tau$ is indeed a bounded stopping time (easy to check) and $$E(X_{\tau})=E(1_AX_n +1_{A^c}X_{n+1})$$

Since the constant $n+1$ is also a bounded stopping time, $E(X_{n+1})=E(X_0)$, thus $$E(1_AX_{n+1})+ E(1_{A^c}X_{n+1}) = E(X_{n+1})=E(X_0) = E(X_\tau)=E(1_AX_n)+ E(1_{A^c}X_{n+1})$$ hence $$E(1_AX_{n+1}) = E(1_AX_n)$$ which concludes the proof.

| cite | improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.