Time Series Decomposition

The relationship between components defines the decomposition structure. An additive model assumes components sum together (y = T + S + R), where seasonal fluctuations remain constant in magnitude regardless of trend level. A multiplicative model assumes components multiply (y = T × S × R), where seasonal amplitude scales proportionally with the trend—common in retail sales where holiday spikes grow as a business expands.

• Additive: Best when seasonal variation is roughly constant over time
• Multiplicative: Best when seasonal variation increases or decreases with trend level
• Diagnostic: Examine if peaks and troughs widen over time to select the correct model

The trend represents the long-term directional movement of the series after removing shorter-term fluctuations. Classical decomposition estimates trend using a centered moving average that smooths out seasonality and noise. For monthly data with annual seasonality, a 12-month moving average eliminates seasonal effects, revealing the underlying growth or decline trajectory.

• Captures structural shifts: market expansion, product maturation, secular decline
• Moving average window must match the seasonal period for clean separation
• Modern approaches use LOESS regression or Hodrick-Prescott filters for adaptive trend estimation

Seasonality captures repeating, fixed-period patterns driven by calendar effects—daily, weekly, monthly, or quarterly cycles. After removing the trend, the seasonal component is computed by averaging values for each period across all cycles. For example, all January values are averaged to produce the January seasonal index.

• Period: The fixed number of observations per cycle (e.g., 7 for daily data with weekly patterns, 12 for monthly data with annual patterns)
• Seasonal indices sum to zero in additive models or average to one in multiplicative models
• Distinct from cyclical patterns, which have variable period lengths tied to economic conditions

The residual (or remainder) is what remains after extracting trend and seasonality—the unexplained, irregular component. A well-decomposed series should produce residuals that resemble white noise: zero mean, constant variance, and no autocorrelation. Residual analysis serves as a diagnostic tool.

• Patterns in residuals indicate incomplete decomposition or missing components
• Outlier detection: Large residual spikes flag anomalous events, promotions, or supply disruptions
• Normality checks: Residual distribution informs prediction interval construction for probabilistic forecasts

Seasonal-Trend decomposition using LOESS (STL) is a robust, modern alternative to classical methods. It handles any seasonal period, adapts to non-linear trends, and is resistant to outliers through iterative weighted local regression. STL is the default choice in production forecasting pipelines due to its flexibility.

• LOESS smoothing: Fits local polynomials to subsets of data, capturing complex trend shapes
• Robust weighting: Down-weights outlier influence across iterations
• Handles missing values natively, unlike classical moving-average approaches
• Supports decomposition of series with seasonal periods longer than annual (e.g., 365 for daily data)

Seasonal adjustment removes the seasonal component to produce a deseasonalized series, enabling clearer trend analysis and serving as a preprocessing step for forecasting. Models like ARIMA operate on the seasonally adjusted series, then seasonal factors are reapplied to generate final predictions.

• Direct forecasting: Model the deseasonalized series, then reseasonalize outputs
• Indirect forecasting: Model the original series with seasonal terms (e.g., SARIMA)
• X-13ARIMA-SEATS: The gold-standard seasonal adjustment procedure developed by the U.S. Census Bureau, incorporating trading day and holiday effects

The long-term directional movement of a time series, representing the underlying growth or decline after removing shorter-term fluctuations.

• Mechanism: Typically extracted using moving averages or polynomial regression.
• Types: Can be deterministic (a fixed function of time) or stochastic (a random walk with drift).
• Business Context: Represents the organic growth of a product category, independent of seasonal promotions or one-off events.
• Example: A steady 2% month-over-month increase in baseline demand for a grocery staple.

A repeating, predictable pattern of fixed frequency within a time series, driven by calendar-related factors.

• Frequency: Can be daily (hourly rush), weekly (weekend dips), or annual (holiday spikes).
• Extraction: Isolated using seasonal differencing or Fourier series.
• Stationarity: Unlike the trend, the seasonal pattern is assumed to have a constant amplitude and period over time.
• Example: A 300% spike in gift wrap sales every December, perfectly repeating each year.

The irregular, unpredictable variation left over after extracting the trend and seasonal components.

• Properties: Should ideally be white noise—a sequence of uncorrelated random variables with a mean of zero.
• Diagnostics: If the residuals exhibit autocorrelation, it indicates that the model has failed to capture all systematic patterns.
• Outlier Detection: Large residual spikes often correspond to anomalous events like stockouts, viral social media posts, or data entry errors.
• Example: A sudden, unexplained dip in sales on a normal Tuesday due to a local power outage.

The mathematical relationship defining how the components interact.

• Additive Model: Y = Trend + Seasonal + Residual. Used when the magnitude of seasonal fluctuations remains constant regardless of the trend level.
• Multiplicative Model: Y = Trend * Seasonal * Residual. Used when seasonal amplitude scales proportionally with the trend (e.g., higher overall sales lead to larger holiday spikes).
• Transformation: A multiplicative model can be converted to an additive one by applying a log transformation to the original series.

Seasonal-Trend decomposition using Loess (Locally Estimated Scatterplot Smoothing). A robust, versatile algorithm for decomposing a time series.

• Advantages: Handles any seasonal frequency, allows the seasonal component to evolve over time, and is robust to outliers.
• Parameters: Controlled by seasonal and trend smoothing windows, which dictate the flexibility of the extracted components.
• Application: Often preferred over classical decomposition for modern, noisy e-commerce data where seasonal patterns shift gradually.
• Example: Modeling a slowly increasing peak in back-to-school shopping that starts earlier each year.

The original, non-parametric method based on moving averages.

• Process: First, a moving average smooths the series to estimate the trend. The trend is removed, and the seasonal component is calculated by averaging the detrended values for each period.
• Limitations: Assumes a fixed, non-evolving seasonal pattern and loses data at the very beginning and end of the series due to the moving average window.
• Use Case: Best suited for stable, mature product lines with highly rigid seasonal cycles and no structural breaks.