Home Statistical Dictionary Martingale Difference Sequence

Martingale Difference Sequence

A sequence of random variables whose conditional expectation given the past is always zero, capturing the idea of 'unpredictable increments' even when the terms are not independent or identically distributed.

Stochastic ProcessesAdvancedFree

In Plain English

A martingale difference sequence formalizes what it means for a sequence of surprises to be genuinely unpredictable: knowing everything up to now gives you no help guessing whether the next term will be positive or negative, its expected value given the past is exactly zero. It's a weaker, more flexible condition than assuming the terms are independent, which is exactly why it shows up throughout time series and econometrics as the standard, realistic assumption for regression errors or forecast errors, terms that don't have to be independent or identically distributed, just unpredictable in this specific conditional-mean sense.

Definition

A sequence of random variables ${ξ_{n}}$ , adapted to a filtration ${F_{n}}$ , is a martingale difference sequence (MDS) if $E [ξ_{n} ∣ F_{n - 1}] = 0$ for every $n$ (and $E ∣ ξ_{n} ∣ < \infty$ ). The name comes directly from its defining role: the partial sums $M_{n} = \sum_{k = 1}^{n} ξ_{k}$ form a martingale, since $E [M_{n} ∣ F_{n - 1}] = M_{n - 1} + E [ξ_{n} ∣ F_{n - 1}] = M_{n - 1}$ . Unlike an i.i.d. sequence, an MDS allows conditional heteroskedasticity (the conditional variance $E [ξ_{n}^{2} ∣ F_{n - 1}]$ can vary with the past) and general dependence in higher moments, only the conditional mean is pinned to zero. A martingale difference central limit theorem, a generalization of the classical CLT, guarantees asymptotic normality of $\frac{1}{n} \sum ξ_{k}$ under mild regularity conditions, without requiring the strict i.i.d. assumption.

Formula

E [ξ_{n} ∣ F_{n - 1}] = 0 for all n

Notation

ξ_{n}

—The $n$-th term of the martingale difference sequence

F_{n - 1}

—The filtration (information available) up to time $n-1$

Properties

Regression and time-series error terms are almost never assumed truly i.i.d. in modern econometric theory, the martingale difference sequence assumption is the realistic middle ground, ruling out predictable bias while still permitting volatility clustering (as in ARCH/GARCH models).
Every i.i.d. sequence with mean zero is automatically a martingale difference sequence, but the reverse is false, MDS is the strictly weaker, more general condition, which is exactly what makes it useful for modeling real, imperfectly independent data.
The martingale difference central limit theorem is one of the most widely invoked asymptotic results in applied econometrics, because it justifies standard normal-approximation inference (t-tests, confidence intervals) even when the underlying errors are heteroskedastic or weakly dependent.

At a Glance

Defining property

E [ξ_{n} ∣ F_{n - 1}] = 0

: conditionally unpredictable, on average

Relation to martingalespartial sums of an MDS are exactly a martingale

Weaker than i.i.d.allows conditional heteroskedasticity and general dependence beyond the mean

Key toolthe martingale difference CLT, asymptotic normality without i.i.d.

Last updated September 8, 2026← Back to the Dictionary