Additive Versus Innovation Outliers in Time Series

An analyst looking at a return series spots one day with a bizarre value and asks the natural question: was this a one-time data glitch that has nothing to do with the underlying process, or was it a real shock that the process actually experienced and is still reacting to? These are the two classic outlier types in time series — additive outliers and innovation outliers — and they look identical on the day they occur but have opposite implications for every day that follows.

An analogy: a scratch on the recording versus a loud noise in the room

Imagine a microphone recording a quiet room. An additive outlier is like a single scratch or pop injected directly onto the recording — a fault of the recording equipment on that one frame, with no bearing on the actual sound in the room before or after. An innovation outlier is like someone actually shouting in the room at that moment — a real event that, because the room has an echo, continues to be audible (fading gradually) in the following few seconds of the recording too. Looking only at the one loud frame, both cases sound identical; the difference only shows up in what happens afterward — silence resuming immediately (additive) versus a fading echo (innovation).

The models, one symbol at a time

Let the "clean" process follow an ARMA model, and let $Y_t$ be what's actually observed. An additive outlier (AO) at time $T$ enters as

$$Y_t = X_t + \omega \, \mathbb{1}\{t = T\},$$

where $X_t$ is the clean underlying ARMA process and $\omega$ is a one-time jolt added only at $t=T$, with no effect on any other time. In plain English: the underlying process is completely unaffected; only the single observed value at time $T$ is corrupted, as if by a recording glitch. An innovation outlier (IO) instead enters the shock that drives the process itself:

$$X_t = \phi_1 X_{t-1} + \dots + \varepsilon_t, \qquad \varepsilon_T = \varepsilon_T^{\text{normal}} + \omega,$$

so the jolt $\omega$ becomes part of the process's own innovation at time $T$, and then propagates forward through the model's usual dynamics (e.g., an AR(1) with $\phi_1=0.7$ would carry roughly 70% of the shock into the next period, 49% into the one after, and so on). In plain English: the shock genuinely happened to the underlying process, so it echoes forward according to the process's own memory, rather than vanishing after one observation.

Worked example 1: distinguishing them from the data

An AR(1) return series with $\phi_1 = 0.6$ shows an unusual value of $+8 on day 50 (typical days are within $±1). If this is an AO, day 51 should look like an ordinary day again, roughly $±1, because the jolt didn't touch the underlying process. If it's an IO, day 51 should show a residual echo of about 0.6 \times 8 = \4.8$above what an ordinary day would show, day 52 about$0.6^2 \times 8 \approx $2.9, and so on, decaying geometrically. Suppose the actual data shows day 51 at \4.5 above baseline and day 52 at $2.6 above baseline — this pattern of decaying carry-over strongly suggests an innovation outlier, not an additive one.

Worked example 2: why the classification changes your response

A risk manager sees a huge one-day loss in a strategy's P&L. Treating it as an AO (assuming it's a data glitch or an isolated fluke) means doing nothing structural — just noting the day and moving on, trusting the model's ongoing forecasts unchanged. Treating it as an IO (a real shock the process absorbed) means updating the forecast for the next several days to include a decaying elevated-risk premium, since the model says the effect hasn't fully dissipated. Formally testing which type fits — using likelihood-ratio-style tests comparing an AO-adjusted fit to an IO-adjusted fit — and getting it backward means either needlessly panicking over a data blip (false IO) or dangerously ignoring a real ongoing shock (false AO).

What this means in practice

Distinguishing additive from innovation outliers matters directly for forecasting and risk management: an AO justifies cleaning the single bad point and otherwise trusting the model unchanged, while an IO means the forecast for several subsequent periods must be adjusted to reflect a real, still-decaying shock. Getting this classification wrong either causes needless data-cleaning of a genuine market event or, worse, silently ignores an ongoing elevated-risk period as if it were noise. This distinction is a standard diagnostic step before robustly refitting an ARIMA model on data suspected to contain outliers.