Must a martingale stay constant?

No. Individual paths may fluctuate considerably.

Is a constant expected value enough?

No. The conditional expectation given current information must match the current value.

Can stopping change the expected result?

It can if the conditions of the relevant stopping theorem are not satisfied; bounded stopping is one important well-behaved case.

AI Glossary / Mathematics and optimization / Intermediate

Martingale (probability theory)

Definition

A martingale is a stochastic process whose next value has conditional expectation equal to its current value, given the information available so far. In discrete time, X_n must be known at time n, have finite expected absolute value, and satisfy E[X_(n+1)|F_n]=X_n. The information sequence F_n, called a filtration, is an essential part of the definition.

What it means

A martingale can move up or down substantially. Its defining feature is the absence of a predictable mean change relative to the specified information, not the absence of uncertainty or risk. The next outcome need not equal the current value on any particular path.
Conditional expectation is stronger than an unchanged overall average. A process may have the same average at every time while its next movement becomes predictable once the current state is known. Whether a process is a martingale therefore depends on both its probability model and what information the observer possesses.

How it works

Identify the process and the information available at each time. Verify that the current value is measurable from that information and has a finite expected absolute value. Then compute the next conditional expectation and compare it with the current value.
One important construction takes an integrable future quantity Z and defines M_n=E[Z|F_n] as information is revealed. Consistent conditional expectations form a martingale. Results about stopping such processes require additional conditions. For example, a bounded stopping time preserves the expectation of an integrable discrete-time martingale; arbitrary unbounded stopping rules cannot be substituted without checking the relevant theorem.

Three practical examples

A probability model can track the evolving conditional estimate of a fixed future outcome as observations arrive.
An algorithm analysis can use a carefully constructed martingale to study randomness revealed during computation.
A simulation study can test a claimed zero-conditional-drift model by examining how its assumptions interact with the recorded information.

Think of it this way

Think of a running estimate that already incorporates everything currently known. New evidence may move it either way, but before seeing that evidence the model expects no average revision. The analogy is limited because real human estimates may be inconsistent; the martingale property is an exact mathematical condition under a specified probability model.

Worked example

Consider an original two-toss experiment using independent fair coins. Define Z=1 if both tosses are heads and Z=0 otherwise. Let M_n be the conditional probability of that event after observing n tosses. Before any observation, M_0=1/4.
After the first toss, M_1=1/2 if the result is heads, because the second toss must still be heads. If the first toss is tails, M_1=0 because the event is already impossible. The expected next estimate is 0.5 times 0.5 plus 0.5 times 0, equaling 0.25=M_0.
After two tosses, M_2 is 1 on HH and 0 on HT, TH, or TT. Conditional on the first toss being heads, its expectation is 0.5, matching M_1 on that branch. Conditional on the first toss being tails, it is certainly zero, also matching M_1. All values are bounded, so integrability is immediate. This verifies the martingale condition branch by branch.
Notice that the realized estimate is not constant. On HH it follows 0.25,0.5,1; on HT it follows 0.25,0.5,0. The martingale property concerns conditional averages, not a promise that individual paths remain near their initial value.
Now examine a different process: X_0=0, X_1 is equally likely to be 1 or -1, and X_2=-X_1. All three unconditional expectations are zero. However, if X_1=1, the next value is known to be -1. Thus E[X_2|X_1=1]=-1, which is not X_1. Constant unconditional expectation is insufficient.
A short stopping example makes another distinction. Let S_n be the running total of +1 for heads and -1 for tails. Stop after the first toss if it is heads; otherwise stop after the second toss. The terminal values are +1 with probability 1/2, 0 with probability 1/4, and -2 with probability 1/4. Their expectation is 0.5+0-0.5=0, equal to S_0. The rule is bounded by two tosses. It does not show that every unlimited stopping or stake-escalation strategy preserves expectation without further assumptions.

Illustrative case study

In a fictional forecasting-methods workshop, River Birch Analytics uses the two-toss model to distinguish conditional and unconditional averages. Participants build a table of what is known at each stage and check the expected revision before revealing the next observation.
The team then reviews a proposed operational forecast model. It treats the martingale property as an assumption to examine, rather than declaring that all real forecasts or market prices must behave that way. This fictional exercise illustrates mathematical reasoning and is not a claim about a deployed forecasting system.

Common misconceptions and mistakes

A martingale is not necessarily an independent sequence, a stationary process, or a low-variance process. Its increments can depend on the past while maintaining conditional mean zero. The information set cannot be omitted: revealing future outcomes can destroy a property that held for an observer who did not know them.
Do not confuse the probability-theory concept with the gambling strategy also called martingale. In particular, unlimited doubling arguments can violate the assumptions required for expectation-preserving stopping results. The mathematical definition does not guarantee a profitable or safe real-world strategy.

Business use

Business relevance lies in analyzing sequential information and stochastic models rigorously. When a report uses a martingale argument, identify the process, filtration, probability assumptions, and theorem conditions. A proof under one model may not survive changing data collection or revealing extra information. In empirical work, approximate evidence of zero drift is also different from a mathematical guarantee of the conditional expectation property.

Check your understanding

Why is X_2=-X_1 not a martingale continuation in the counterexample?

Answer: Once X_1 is known, X_2 is known to have the opposite sign, so its conditional expectation differs from the current value.

Related terms

Conditional expectation
Stochastic process
Stopping time
Filtration
Adapted process
Optional stopping
Martingale difference
Submartingale
Supermartingale

Learn next

Repeat the conditional-probability construction for the event that two independent fair tosses have the same outcome. Compute the estimate before any toss, after the first, and after the second. Verify the conditional-expectation property despite the different path values.

Sources

Primary teaching source: Scott Sheffield, MIT 18.600 Lecture 34, Martingales and the Optional Stopping Theorem: https://ocw.mit.edu/courses/18-600-probability-and-random-variables-fall-2019/3ce8cee748c630414deebc968bd375e1_MIT18_600F19_lec34.pdf . Branch tables, counterexample, and bounded stopping calculation are original.

Previous: Markov reward process Next: Masked autoregressive flow