Structural Time Series Models

Show a monthly sales chart to a manager and they instinctively read it in parts: "there's a steady upward trend, a big Christmas spike every year, and a lot of week-to-week noise." That decomposition into meaningful pieces is exactly how humans understand time series — and it is exactly what an ARIMA model throws away. ARIMA differences the data until it is stationary and fits abstract autoregressive and moving-average coefficients: powerful, but a black box. Ask it "what is the trend?" or "how big is the seasonal effect?" and it cannot answer, because it never represented those things.

A structural time series model (also called an unobserved components model) takes the opposite stance: build the model out of the interpretable pieces themselves. Write the series as a sum of a level, a trend, a seasonal, maybe a cycle, plus noise — each an evolving component with its own equation — and let the data estimate all of them at once. It is a state-space model whose state vector is literally "the current level, slope, and seasonal factors", which means the whole thing is fitted and forecast by the Kalman filter.

The components

Decompose the observation into additive pieces:

y_t = \underbrace{\mu_t}_{\text{level}} + \underbrace{\nu_t}_{\text{seasonal}} + \underbrace{\psi_t}_{\text{cycle}} + \underbrace{\varepsilon_t}_{\text{noise}},

and give each piece a small state equation describing how it drifts over time. The most common building blocks:

Local level. \mu_t = \mu_{t-1} + \eta_t — a randomly wandering baseline (a random walk).
Local linear trend. Add a slope that itself drifts: \mu_t = \mu_{t-1} + b_{t-1} + \eta_t, b_t = b_{t-1} + \zeta_t — a level whose rate of change evolves.
Stochastic seasonal. Seasonal factors that sum to (about) zero over a year but are allowed to change shape slowly, so the seasonal pattern can evolve rather than repeat identically forever.
Cycle. A stochastic quasi-periodic term for business cycles — a longer, less regular swing than the fixed calendar season.

Crucially, each component has its own variance controlling how fast it evolves. Set a variance to zero and that component becomes deterministic — a fixed slope, an unchanging season. Let it be positive and the component adapts over time. Those variances are the knobs you estimate from the data.

Components adding up to the series

Below, three curves. The straight-ish rising line is the level + trend \mu_t. The wave near the bottom (drawn around its own baseline for clarity) is the seasonal component \nu_t. The top jagged line is the observed series — the sum of the two, plus a little noise. The structural model's job is the reverse: given only the observed line, use the Kalman filter to peel it apart into these interpretable pieces.

Because each piece is a genuine part of the model, you can plot the estimated trend on its own, report the seasonal effect for each month with an uncertainty band, or ask how the slope has changed over a decade — questions an ARIMA fit simply cannot answer.

Why authors reach for structural models

Interpretability. Every parameter and every state component means something — a level, a growth rate, a December effect. You can hand the decomposition to a non-statistician and they will understand it.
Missing data and irregular spacing. Inherited free from the state-space machinery — skip the update on a gap and carry the components forward. No differencing headaches.
Evolving components. The trend and the season are allowed to change over time, so the model adapts to a series whose behaviour shifts — where a fixed-coefficient ARIMA would lag.
Easy to extend. Bolt on a regression term (holidays, a price variable, an intervention dummy) simply by adding it to the observation equation — the filter handles the rest.

This flexibility is why structural models power modern Bayesian structural time series (BSTS) and nowcasting — combining many noisy, mixed-frequency indicators (search trends, card transactions, surveys) into a real-time estimate of a slow official statistic like GDP. Google's CausalImpact, and much of the "measure the effect of an intervention on a time series" toolkit, is built directly on structural state-space models.

Structural vs ARIMA — two philosophies

The two are not enemies; many structural models have an equivalent ARIMA representation (the local-level model is \text{ARIMA}(0,1,1), the local linear trend is \text{ARIMA}(0,2,2)). The difference is one of emphasis:

Question	ARIMA	Structural
Built from	differences + AR/MA lags	interpretable components
"What is the trend?"	not represented	a named state component
Handling nonstationarity	differencing	a random-walk component
Missing data	awkward	natural (skip the update)
Fitted by	likelihood (often via Kalman)	Kalman filter likelihood

Choose ARIMA when you want a parsimonious black-box forecaster and don't care about the parts; choose a structural model when interpretability, missing data, evolving structure, or regression effects matter.

The seductive thing about structural models is how easily you can keep adding components — a stochastic level, a stochastic slope, a stochastic season, a cycle, an irregular — each with its own variance. The danger is identifiability: with several noise variances all trying to explain the same wiggles, the likelihood can become nearly flat in some directions. The optimiser can't tell whether a given bump belongs to the trend, the cycle or the irregular, and different software (or different starting values) may return quite different fits with almost the same likelihood.

Symptoms are a variance estimate that collapses to exactly zero (the "pile-up" problem — the component is being switched off), wildly wide confidence intervals on the variances, or fits that jump around with the starting point. The cures: keep the model as simple as the data can support, fix components to deterministic (variance zero) unless the data clearly demands they evolve, use a decent length of data, and lean on a Bayesian prior to regularise the variances. More components is not more insight if the data can't distinguish them.

The structural / unobserved-components program is associated above all with Andrew Harvey, whose 1989 book Forecasting, Structural Time Series Models and the Kalman Filter made the case that a model should be assembled from parts you can name and defend, not from differences chosen to pass a stationarity test. It was, in part, a friendly rebellion against the then-dominant Box–Jenkins ARIMA orthodoxy — a debate about whether a model should be a transparent story about level, trend and season, or a compact black box tuned only to forecast. Modern practice happily uses both; but every time you see a chart neatly split into "trend", "seasonal" and "remainder", you are looking at the structural way of seeing.