Martingales and Optional Stopping Intuition

How filtrations track growing information, why martingales formalize conditional fairness, and how optional stopping explains when path-dependent stopping still preserves expectation.
Modified

April 26, 2026

Keywords

martingale, filtration, stopping time, optional stopping, conditional expectation

1 Role

This is the second page of the Stochastic Processes module.

Its job is to introduce the main language of conditional fairness:

  • information grows over time
  • random variables are judged relative to that information
  • stopping rules can themselves depend on the path

This is the page where conditional expectation turns into a structural object rather than only a computation rule.

2 First-Pass Promise

You can read this page on its own inside the full module spine.

If you stop here, you should still understand:

  • what a filtration means at first pass
  • what makes a process a martingale
  • what a stopping time is
  • why optional stopping is powerful but not magic

3 Why It Matters

Martingales appear whenever we need to formalize:

  • fair games
  • no predictable gain from current information
  • path-dependent stopping rules
  • sequential uncertainty accumulation

That is why martingale language keeps reappearing in:

  • concentration arguments
  • hitting-time questions
  • gambling and finance intuition
  • stochastic calculus
  • RL and sequential decision theory

The point is not only that martingales are elegant.

They give a clean way to say:

given everything known so far, the process has no systematic expected upward drift

4 Prerequisite Recall

  • from Probability, conditional expectation means best expectation given current information
  • from Markov Chains and Stationary Distributions, stochastic processes evolve over time and need a notion of what is known at each step
  • from random-walk intuition nearby, summing fair mean-zero increments creates a natural candidate martingale

5 Intuition

5.1 A Filtration Is Growing Information

At first pass, a filtration

\[ \mathcal{F}_0 \subseteq \mathcal{F}_1 \subseteq \mathcal{F}_2 \subseteq \cdots \]

is just the record of what has been revealed by time n.

So:

  • \mathcal{F}_0 is what you know at the start
  • \mathcal{F}_1 includes one more step of information
  • later sigma-fields include everything earlier plus new observations

This is the bookkeeping device that makes “known so far” precise.

5.2 A Martingale Means Conditional Fairness

A process (M_n) is a martingale if its current value is already the best prediction of its next value given present information.

At first pass, read that as:

no predictable expected gain remains once the present is known

That does not mean the path is flat.

It can fluctuate a lot.

It only means those fluctuations do not create systematic conditional drift.

5.3 A Stopping Time Must Use Present And Past Information Only

A stopping time \tau is a random time whose decision rule can be made from information already revealed by that time.

So you may stop:

  • when wealth first hits a threshold
  • when a process exits an interval
  • when a queue first empties

But you may not stop using future information you have not observed yet.

5.4 Optional Stopping Is About When Fairness Survives Stopping

The naive gambler’s fantasy is:

if I wait until a good time to stop, maybe I can force a positive expected gain

Optional stopping says that for a martingale, this does not work automatically.

Under safe conditions such as bounded stopping times, stopping does not create expected profit out of nowhere.

That is why optional stopping is powerful:

  • it analyzes path-dependent stopping rules
  • but it also blocks fake arbitrage stories

6 Formal Core

Definition 1 (Definition: Filtration) A filtration ( \mathcal{F}_n ) is an increasing sequence of sigma-fields:

\[ \mathcal{F}_0 \subseteq \mathcal{F}_1 \subseteq \mathcal{F}_2 \subseteq \cdots \]

At first pass, \mathcal{F}_n means the information available by time n.

Definition 2 (Definition: Martingale) A process (M_n) adapted to ( \mathcal{F}_n ) is a martingale if:

  1. M_n is integrable
  2. M_n is \mathcal{F}_n-measurable
  3. for every n,

\[ \mathbb{E}[M_{n+1} \mid \mathcal{F}_n] = M_n \]

This is the formal statement of conditional fairness.

Definition 3 (Definition: Stopping Time) A random time \tau is a stopping time with respect to ( \mathcal{F}_n ) if the event { \tau \le n } can be determined using information in \mathcal{F}_n.

At first pass, this means the rule “stop now” never needs future knowledge.

Theorem 1 (Theorem Idea: Optional Stopping Under Safe Conditions) If (M_n) is a martingale and \tau is a bounded stopping time, then

\[ \mathbb{E}[M_{\tau}] = \mathbb{E}[M_0]. \]

This is the first safe version of optional stopping.

The moral is:

  • path-dependent stopping can still preserve fairness
  • but only under conditions that prevent pathological infinite waiting or uncontrolled growth

7 Worked Example

Let (X_1, X_2, \dots) be independent fair coin increments taking values +1 and -1 with probability 1/2, and define

\[ M_n = X_1 + \cdots + X_n. \]

Then (M_n) is a martingale.

Now define the stopping time

\[ \tau = \min\{n : M_n = 1\} \wedge 2. \]

This means:

  • stop immediately if the walk hits 1
  • otherwise stop by time 2 no matter what

Possible outcomes:

  • with probability 1/2, X_1=1, so \tau=1 and M_{\tau}=1
  • with probability 1/4, X_1=-1, X_2=1, so \tau=2 and M_{\tau}=0
  • with probability 1/4, X_1=-1, X_2=-1, so \tau=2 and M_{\tau}=-2

Therefore

\[ \mathbb{E}[M_{\tau}] = \frac12(1) + \frac14(0) + \frac14(-2) = 0 = \mathbb{E}[M_0]. \]

So even though the stopping rule depends on the revealed path, bounded optional stopping preserves the martingale expectation.

8 Computation Lens

When you meet a martingale argument, ask:

  1. what is the information filtration?
  2. what random variable or process is being conditioned on that information?
  3. where is the zero conditional drift coming from?
  4. is the stopping rule adapted, or does it secretly use future information?
  5. are the conditions for optional stopping actually safe enough?

Those questions usually reveal whether the proof is really about:

  • fairness
  • pathwise stopping
  • concentration
  • or turning a complicated process into a conditional-expectation invariant

9 Application Lens

9.1 Hitting Times And Sequential Questions

Optional stopping is one of the cleanest tools for analyzing “stop when something first happens” events.

9.2 Concentration And Probability Theory

Many concentration inequalities and maximal inequalities are built on martingale structure.

9.3 Stochastic Calculus And Finance

Continuous-time martingales become part of the language behind Brownian motion, Ito calculus, and no-arbitrage intuition.

10 Stop Here For First Pass

If you stop here, retain these five ideas:

  • a filtration is the formal record of growing information
  • a martingale has no predictable conditional drift
  • a stopping time must be chosen using only present and past information
  • optional stopping says stopping does not automatically create expected profit
  • martingale language is a reusable bridge into concentration, hitting times, and stochastic calculus

11 Go Deeper

The strongest adjacent live pages right now are:

This page closes the full first-pass spine of the module.

12 Optional Deeper Reading After First Pass

The strongest current references connected to this page are:

  • MIT 18.445 lecture 15 - official MIT note introducing conditional expectation and martingales. Checked 2026-04-26.
  • MIT 18.445 lecture 16 - official MIT note focused directly on optional stopping. Checked 2026-04-26.
  • MIT 15.070J lecture 10 - official MIT note on martingales and stopping times. Checked 2026-04-26.
  • Stanford Stats 310C syllabus - official Stanford syllabus showing the route through filtrations, stopping times, martingales, and later Ito calculus. Checked 2026-04-26.

13 Sources and Further Reading

  • MIT 18.445 lecture notes page - First pass - official MIT lecture hub covering martingales and optional stopping in the main stochastic-processes sequence. Checked 2026-04-26.
  • MIT 18.445 lecture 15 - First pass - official MIT note introducing conditional expectation and martingales. Checked 2026-04-26.
  • MIT 18.445 lecture 16 - First pass - official MIT note on optional stopping. Checked 2026-04-26.
  • MIT 15.070J lecture notes page - Second pass - official MIT lecture hub with filtration, martingale, stopping-time, and convergence notes. Checked 2026-04-26.
  • MIT 15.070J lecture 10 - Second pass - useful official MIT note on martingales and stopping times. Checked 2026-04-26.
  • Stanford Stats 218 - Bridge outward - official Stanford stochastic-processes course page where martingales are part of the live curriculum. Checked 2026-04-26.
  • Stanford Stats 310C syllabus - Bridge outward - useful Stanford syllabus showing the progression from filtrations and stopping times toward Brownian motion and Ito calculus. Checked 2026-04-26.
Back to top