Classical Decomposition

Classical decomposition breaks a time series into trend, seasonality, and residual noise so the remaining series can be modeled with stationary-process tools.

1. Additive Model

$$X_t = m_t + s_t + Y_t$$
ComponentDescriptionProperties
$m_t$Deterministic trende.g., linear $a + bt$, polynomial $\sum_{k=0}^p a_k t^k$
$s_t$Deterministic seasonal componentPeriodic: $s_t = s_{t+d}$ with period $d$
$Y_t$Random (noise) processZero-mean, often modeled as Gaussian

When seasonality is recorded over one full period, it is often normalized by

$$\sum_{j=1}^d s_j = 0$$

so the average level is not counted twice in both $m_t$ and $s_t$.

2. Multiplicative View

If seasonal amplitude grows with the level of the series, a multiplicative structure is more natural. A common workaround is to log-transform:

$$\log(X_t) = m_t + s_t + Y_t$$

which converts multiplicative behavior into an additive decomposition.

3. Trend Estimation

Polynomial Regression

Fit $\hat{m}_t = \hat{a} + \hat{b} t$ (or higher degree) by minimizing

$$\sum_{t=1}^n (X_t - a - bt)^2$$

Moving Average Filter

$$\hat{m}_t = \frac{X_{t-q} + \cdots + X_t + \cdots + X_{t+q}}{2q+1}, \quad q+1 \leq t \leq n-q$$

Limitations:

  1. All past observations have equal weight (but recent observations usually matter more)
  2. Cannot estimate $m_t$ near the endpoints — cannot be used for Forecasting

4. Trend Elimination

Use the difference operator $\nabla X_t = X_t - X_{t-1}$.

If $m_t$ is a polynomial of degree $p$, then $p$ applications of $\nabla$ eliminate the trend:

$$\nabla^p m_t = \text{constant}$$

5. Seasonality Elimination

Use the lag-$d$ difference operator:

$$\nabla_d X_t = X_t - X_{t-d} = (1 - B^d)X_t$$

If $s_t$ has period $d$ (i.e., $s_t = s_{t+d}$), then $\nabla_d$ eliminates $s_t$:

$$\nabla_d s_t = s_t - s_{t-d} = 0$$

For multiple seasonal components with periods $d_1, \dots, d_k$, the smallest $d$ such that $\nabla_d$ eliminates all seasonality is $d = \text{lcm}(d_1, \dots, d_k)$.

6. Course Strategy

In STA457, decomposition is usually a preprocessing step:

  1. Identify trend and seasonal structure.
  2. Remove them using regression, smoothing, or differencing.
  3. Treat the residual component as approximately stationary.
  4. Fit AR, MA, or ARMA structure to the residual series.