Notes
🔍

Additive Decomposition

Some time series feel like they’re made of a few simple “layers” stacked on top of each other: a long-term direction, a repeating seasonal pattern, and some random wiggles. Additive decomposition is the idea of peeling those layers apart so you can see (and model) each one more clearly.

What it means (the model)
In an additive decomposition, the observed value at time t is written as a sum of components:

y_t = T_t + S_t + R_t

where T is the trend (slow movement over time), S is the seasonal component (repeating pattern like weekly or yearly cycles), and R is the remainder/residual (what’s left: noise, shocks, unexplained variation). “Additive” specifically means the seasonal ups and downs have roughly the same size even when the series level changes.

When additive is a good fit
Additive decomposition tends to work well when seasonal variation is fairly constant in absolute terms:

  • Monthly electricity usage that rises by ~200 kWh every summer, regardless of the year’s baseline.
  • Store foot traffic that adds ~50 extra visits on weekends even as the business grows.
  • Exam scores with a steady improvement trend plus a consistent “end-of-term” bump.

Why it matters in AI/ML
Many forecasting workflows improve when you separate structure from noise. Decomposition helps:

  • Build features: use T and S as inputs to regression, gradient boosting, or neural nets.
  • Stabilize modeling: train models on the remainder after removing seasonality/trend.
  • Diagnose problems: leftover patterns in R can reveal missing seasonality, change points, or anomalies.

You’ll see additive decomposition in tools like STL (Seasonal-Trend decomposition using Loess) and in libraries such as statsmodels and Prophet (which uses an additive formulation by default).

Additive Decomposition models a time series as the sum of components, typically level/trend + seasonality + irregular noise (and sometimes a cycle): y(t)=T(t)+S(t)+E(t). It assumes seasonal fluctuations have roughly constant magnitude over time, unlike multiplicative models. It is important for isolating structure from noise, enabling forecasting and anomaly detection. Example: decomposing daily website visits into a long-term growth trend plus a weekly pattern and residual spikes.

Imagine your monthly electricity bill is made up of a few simple parts: a basic amount you pay most months, extra cost in winter or summer, and some random surprises. Additive decomposition treats a time-based data line (a time series) the same way: it says the value at each time point can be written as a sum of separate pieces.

Typically those pieces are a long-term trend (overall up or down), a repeating seasonal pattern (like weekends or holidays), and leftover noise (random wiggles). It’s called “additive” because these parts add together, and their sizes stay roughly consistent over time.