Time Series Components

Why This Matters

Time series analysis is the backbone of forecasting. Your ability to decompose data into its fundamental components determines how well you can model real-world phenomena. The core skill here is distinguishing systematic patterns from random variation, isolating each component, and knowing when to apply specific techniques to make messy data analyzable.

These concepts connect directly to regression, ARIMA modeling, and forecast evaluation, so everything that comes later builds on what's covered here. When you see a time series, you should immediately ask: What's the trend? Is there seasonality? Is this stationary?

---

Systematic Patterns: The Predictable Structure

Systematic components are the portions of a time series that follow recognizable, modelable patterns. These are what forecasting models try to capture. Everything left over is noise.

Trend

Trend is the long-term directional movement in the data: upward (growth), downward (decline), or flat (stability).

• Trends can be linear, following something like $y = \beta_0 + \beta_1 t$, or nonlinear, following polynomial or exponential forms.
• Extrapolating trend is often the primary goal of time series analysis, which is why correctly identifying trend shape matters so much for forecasting accuracy.

Seasonality

Seasonality refers to fixed-period, recurring patterns that repeat at known intervals (daily, weekly, monthly, quarterly, or annually).

• These patterns are driven by external factors like weather, holidays, school calendars, or fiscal year cycles.
• The defining feature is that the period is constant and known in advance. Retail sales spiking every December is a textbook example.

Cyclical Patterns

Cyclical patterns are longer-term fluctuations tied to business or economic cycles, typically spanning multiple years.

• Unlike seasonality, cycles have a variable and unpredictable period. You can't say "this repeats every X months."
• They're harder to model precisely because they don't follow a fixed schedule and may require leading economic indicators to forecast.

Level

Level is the baseline value around which the series fluctuates. Think of it as the "center of gravity" of your data.

• It serves as the reference point for decomposition: trend describes movement away from level, and seasonality describes oscillation around it.
• A sudden shift in level (a structural break) can indicate a regime change, like a new policy or market disruption, that requires model adjustment.

---

Random Variation: The Unpredictable Noise

Not everything in a time series can be explained by patterns. Understanding what's random, and confirming that your model residuals look random, is essential for valid inference.

Irregular Fluctuations

Irregular fluctuations are unpredictable, non-repeating variations caused by one-time events like natural disasters, policy shocks, or market surprises.

• After you remove the systematic components, what's left are these fluctuations (also called residuals or noise).
• They can't be forecasted, but they still matter: you need to account for them in prediction intervals and model diagnostics.

White Noise

White noise is a purely random series with three specific properties: $\text{mean} = 0$, constant variance, and zero autocorrelation at all lags.

• Formally: $\epsilon_t \sim WN(0, \sigma^2)$, where observations are uncorrelated.
• White noise is the benchmark for model adequacy. If your residuals aren't white noise, your model missed something systematic.

---

Statistical Properties: What Makes Analysis Possible

These concepts determine whether standard time series methods will actually work on your data. Stationarity and autocorrelation are diagnostic checkpoints before any serious modeling begins.

Stationarity

A series is stationary when its statistical properties (mean, variance) stay constant over time, regardless of when you observe it.

• Many models, including ARMA, require stationarity. Non-stationary data violates their assumptions and can produce spurious results.
• You can often achieve stationarity through transformation. Common approaches:
  • Differencing: $y_t - y_{t-1}$ removes trend
  • Log transform: stabilizes increasing variance
  • Detrending: subtracting a fitted trend line

Autocorrelation

Autocorrelation measures the correlation of a series with its own lagged values. It tells you how today's value relates to yesterday's, last week's, and so on.

• The ACF (autocorrelation function) quantifies this at each lag $k$: $\rho_k = \text{Corr}(y_t, y_{t-k})$
• ACF and PACF (partial autocorrelation function) plots are your primary tools for model selection. The patterns in these plots reveal whether AR, MA, or ARMA structures are appropriate.

---

Analytical Techniques: Tools for Understanding

These methods help you extract insights from raw time series data. Decomposition separates the signal; smoothing reveals it.

Decomposition

Decomposition separates a time series into trend, seasonality, and residuals to reveal underlying structure hidden in the raw data.

There are two main forms:

• Additive model: $y_t = T_t + S_t + \epsilon_t$. Use this when seasonal swings stay roughly the same size regardless of the level.
• Multiplicative model: $y_t = T_t \times S_t \times \epsilon_t$. Use this when seasonal swings grow proportionally with the level.

A good rule of thumb: if a plot shows December spikes getting larger as overall sales grow, that's multiplicative. If the spikes stay about the same height year over year, that's additive.

Moving Average

A moving average is a smoothing technique that averages consecutive observations to reduce noise and reveal trend.

• Simple MA weights all $k$ observations equally: $MA_t = \frac{1}{k}\sum_{i=0}^{k-1} y_{t-i}$
• Weighted MA gives more emphasis to recent values.
• There's a key tradeoff in choosing window size: larger $k$ produces a smoother trend but introduces more lag; smaller $k$ is more responsive but noisier.

---

Quick Reference Table

---

Self-Check Questions

1. A retail company notices sales increase every December but also sees larger overall swings during economic expansions. Which two components explain these patterns, and how do they differ in predictability?

2. You've fit a forecasting model and want to check if it captured all systematic patterns. What statistical property should the residuals exhibit, and how would you test for it?

3. Compare and contrast additive versus multiplicative decomposition. Under what data conditions would you choose one over the other?

4. Your time series has an upward trend and increasing variance over time. Which two techniques might you apply to achieve stationarity, and in what order?

5. If you're asked to "identify the appropriate model structure" for a dataset, which diagnostic tool would you use to examine dependence across time lags, and what patterns would suggest an AR versus MA component?