The Wold Decomposition

Why does it make sense to model so many different stationary time series — stock returns, interest rate spreads, inflation surprises — using the same basic toolkit of autoregressive and moving-average terms? It isn't a coincidence or a convenient simplification; it's a mathematical guarantee. The Wold decomposition theorem says that (almost) any stationary time series, no matter how it was actually generated, can be rewritten as a sum of a perfectly predictable component and an infinite weighted sum of past random shocks. This is why ARMA models aren't just one modeling choice among many — they're the natural language for describing stationary processes in general.

An analogy: an echo chamber

Think of a room with a sound source and echoes. What you hear at any moment is a blend of today's fresh sound plus faded, decaying echoes of everything said before — an echo from a second ago is loud, an echo from a minute ago has faded to almost nothing, but in principle every past sound still contributes something, however small. The Wold decomposition says a stationary time series works exactly like this room: today's value is a mix of a brand-new random shock plus decaying echoes of every past shock, weighted so that older shocks matter less and less. Almost nothing about the original mechanism generating the sounds matters for this description — any stationary process, whatever produced it, can be re-expressed this way.

The theorem, one piece at a time

Any stationary process $X_t$ (excluding a technical class of purely deterministic ones) can be written as:

$$X_t = \sum_{j=0}^{\infty} \psi_j \varepsilon_{t-j} + \kappa_t, \qquad \varepsilon_t \sim \text{white noise},$$

where $\varepsilon_t$ is a sequence of uncorrelated shocks (white noise — the fresh, unpredictable news arriving each period), $\psi_j$ are weights describing how much a shock from $j$ periods ago still influences today (with $\psi_0=1$ and $\sum \psi_j^2 < \infty$, meaning the weights shrink fast enough for the sum to make sense), and $\kappa_t$ is a deterministic component — perfectly predictable from the past, like a fixed cycle or a component fully explained by prior values with no residual randomness, which is usually zero or negligible for financial series. In plain English: any stationary series is (almost) always expressible as a weighted, decaying sum of all its own past surprises, plus at most a perfectly forecastable leftover — this is exactly the structure an infinite-order moving average (MA(∞)) has, which is why finite ARMA models, which are just convenient truncated/rational approximations to this infinite sum, work so broadly.

Worked example 1: an AR(1) process is secretly an MA(∞)

Take the AR(1) process $X_t = 0.5 X_{t-1} + \varepsilon_t$. Substituting repeatedly: $X_t = \varepsilon_t + 0.5\varepsilon_{t-1} + 0.25\varepsilon_{t-2} + 0.125\varepsilon_{t-3} + \cdots$, i.e. $\psi_j = 0.5^j$. This is exactly the Wold form: today's value is the newest shock, plus 50% of last period's shock, plus 25% of the shock from two periods ago, and so on, geometrically decaying. A shock five periods back still contributes $0.5^5 = 0.03125$, about 3% of its original size — small but, in principle, never fully zero.

Worked example 2: reading persistence off the weights

Two return series both have an immediate reaction to news of the same size ($\psi_0=1$), but series A has weights $\psi_j = 0.3^j$ and series B has weights $\psi_j = 0.8^j$. After 3 periods, A's initial shock has decayed to $0.3^3=0.027$ (2.7% of its original size) while B's has decayed only to $0.8^3=0.512$ (51.2%). Series A's shocks fade almost immediately — news is absorbed fast — while series B carries the memory of a shock for much longer, meaning today's value is still meaningfully explained by news from several periods back. This is exactly the intuition behind measuring persistence via an estimated MA representation rather than eyeballing the raw series.

Watch a mean-reverting path and notice how each new random shock's influence visibly fades over subsequent steps rather than vanishing instantly — that fading is the $\psi_j$ weight sequence made visible.

What this means in practice

The Wold decomposition is why fitting an ARMA model to a stationary financial series isn't an arbitrary modeling choice — it's the natural finite approximation to a representation that (almost) any such series has. It also underlies impulse-response analysis: to ask "how does a shock today propagate into the future," you're asking about the $\psi_j$ weights, whether the series comes from an AR, MA, ARMA, or something else entirely.