Time-series decomposition and forecasting

Description

Separates a metric's history into trend, seasonal, cyclical and irregular parts, and forecasts it by smoothing level, trend and season or by regression on its own lags.

A time series is modelled as trend, seasonal, cyclical and irregular components, combined additively or multiplicatively. Centred moving averages over one seasonal cycle (4 quarters, 12 months, 7 days) remove seasonality and noise to reveal trend. The seasonal index method expresses each observation as a percentage of its centred moving average, takes the median per season, rescales to average 100, and divides the series by these indices to seasonally adjust it; X-11 methods allow seasonality to evolve. For forecasting, simple exponential smoothing weights the newest observation against the previous level with constant alpha; Holt-Winters adds trend (beta) and seasonal (gamma) recursions. Autoregressive models regress the series on its own lags, choosing order by testing down from a maximum; ARIMA extends this through iterative identification, estimation and checking. All methods assume past patterns continue, so structural breaks must be watched for.

Attributes

Loss type
Siloed
Effort
Days

Steps

  1. Plot the series and judge whether trend, seasonality, cycles and breaks are present.
  2. Compute centred moving averages over one seasonal period to estimate the trend-cycle.
  3. Express each observation as a percentage of its centred moving average, take the median per season, rescale to average 100, and divide the series by these seasonal indices.
  4. Check the adjusted series for leftover seasonality; if the pattern evolves, use X-11 or STL-type decomposition.
  5. To forecast, choose simple exponential smoothing (no trend or season), Holt-Winters (trend and season), or an AR or ARIMA model, setting smoothing constants or lag order by minimising one-step forecast errors.
  6. Generate forecasts with intervals, compare them with holdout periods, and review them against known launches, campaigns or breaks.

Sources

  • Statistics for Business and Economics (Newbold et al., 2022)
  • Statistics for Business: Decision Making and Analysis (Stine & Foster, 2018)
  • Introduction to Nonparametric Regression (Takezawa, 2006)
  • Applied Bayesian Modelling (Congdon, 2003)
  • The Analysis of Time Series: An Introduction (Chatfield, 2003)
  • Analysis of Financial Time Series (Tsay, 2002)
  • AI-for-quantitative-UX-research (Rafiei & Jozranjbar, GitHub repository, 2026)
  • Quantitative Social Research Methods (Singh, 2007)
  • Bayesian Statistical Modelling (Congdon, 2006)
  • Customer Analytics for Dummies (Sauro, 2015)

Also known as

  • Time-series decomposition and seasonal adjustment
  • Classical decomposition
  • Seasonal index method
  • Ratio-to-moving-average
  • Seasonal adjustment
  • Centred moving average
  • Exponential smoothing
  • Holt-Winters
  • Triple exponential smoothing
  • Autoregressive model AR(p)
  • ARIMA
  • Box-Jenkins forecasting

Categories

  • Analytics & Behaviour
  • Statistics & Modelling