Lesson 1 - Trend, Seasonality, and Residual

Welcome to Trend, Seasonality, and Residual#

Module 1 called Cyclepath’s shape “an upward trend plus yearly seasonality” without pinning down what those words mean precisely. Now you will. A time series is rarely one signal — it’s usually two or three signals added or multiplied together, and decomposition is the act of pulling them back apart. Once you can see trend, seasonality, and residual as separate, nameable things, half the rest of this course is just deciding what to do with each piece on its own.

By the end of this lesson, you will be able to:

• Define trend, seasonality, and residual, and tell them apart on sight
• State the additive and multiplicative decomposition equations and when each applies
• Explain why decomposition comes before modeling, not after
• Preview how this module will build a decomposition by hand and then with a library

Let’s name the three pieces.

The Three Components#

Every classical decomposition splits a series y into three components:

• Trend (T) — the slow-moving, long-run direction of the series, with the seasonal wiggle and the noise smoothed away. Cyclepath’s trend is a steady climb as the bike network grows; on other series a trend might be flat, falling, or curved.
• Seasonality (S) — a pattern that repeats on a fixed, known calendar cycle: every 12 months for Cyclepath, every 7 days for a series with a weekly pattern, every 24 hours for one with a daily pattern. Seasonality is not “any repeating wiggle” — it’s specifically tied to the calendar, which is what makes it predictable rather than merely observed.
• Residual (R), also called the remainder or irregular component — everything left over after trend and seasonality are removed. In a good decomposition the residual looks like unstructured noise: no visible pattern, no trend, no seasonality of its own. If the residual still has a shape to it, the decomposition missed something.

Notice the checks built into that picture: the trend panel has no wiggle left in it, the seasonal panel has no drift left in it, and the residual panel has neither — just noise. That’s what a decomposition is for: each component should contain only the kind of structure its name promises, and nothing else.

How the Pieces Combine#

There are two standard ways to recombine the three components back into the observed series, and the difference matters more than it looks:

• Additive: y = T + S + R. The seasonal swing is a constant amount — Cyclepath’s summer peak is roughly 3,200 trips above the trend line every year, whether the trend is at 10,000 or 16,000.
• Multiplicative: y = T × S × R. The seasonal swing is a constant percentage of the trend — a series where the summer peak is always “about 25% above the current level,” so the absolute swing grows as the level grows.

Both are legitimate models; which one fits depends entirely on whether the seasonal swing you observe stays a fixed size or scales with the level. Guessing wrong doesn’t just look slightly off — it produces a residual that still has a pattern in it (usually one that grows or shrinks over time), which defeats the entire point of decomposing in the first place. Lesson 3 makes this choice rigorously, on real data, with real evidence.

What’s Ahead in This Module#

This module builds decomposition in three passes, each more sophisticated than the last:

1. By hand (Lesson 2) — a centered moving average estimates the trend, subtracting it exposes the seasonal shape, and what’s left is the residual. No library required; you’ll see exactly what a decomposition function does internally.
2. Additive vs. multiplicative, formalized (Lesson 3) — a rigorous test for which combination rule fits a given series, applied to Cyclepath and to a contrasting example where the swing genuinely grows.
3. STL (Lesson 4) — a more robust decomposition method that handles outliers and slowly changing seasonality better than the classical moving-average approach, and is what most practitioners reach for today.

The module capstone (Lesson 5) applies all three to the full Cyclepath series and interprets what each component reveals.

Practice Exercises#

Exercise 1: Name the component#

A retailer’s daily sales series has a steady rise every December (holiday shopping) that repeats every year at the same time, on top of a gentle year-over-year growth as the store gains customers. Which part is trend and which is seasonality?

Exercise 2: Spot the bad residual#

You decompose a series and plot the residual. Instead of unstructured noise, it clearly still rises and falls every 12 months, just with smaller swings than the original series. What does that tell you?

Exercise 3: Additive or multiplicative?#

Two series both have a rising trend and a yearly peak. Series A’s peak is 500 units above trend every year, no matter how high the trend has climbed. Series B’s peak is always about 20% above whatever the trend currently is. Which is additive and which is multiplicative?

Summary#

Decomposition splits an observed series y into three named components. Trend (T) is the slow-moving long-run direction with seasonality and noise smoothed away. Seasonality (S) is a pattern tied to a fixed calendar cycle — not just any repeating wiggle. Residual (R) is whatever is left over, and in a good decomposition it looks like unstructured noise. The pieces recombine either additively (y = T + S + R, constant-size seasonal swing) or multiplicatively (y = T × S × R, seasonal swing scales with the level) — picking the wrong one leaves a patterned residual, which is exactly the tell that something’s off. This module builds a decomposition by hand, formalizes the additive-vs-multiplicative choice, and introduces STL, before applying all three to the full Cyclepath series.

Key Concepts#

• Trend — the slow-moving long-run direction, with seasonality and noise removed.
• Seasonality — a pattern repeating on a fixed calendar cycle, not any repeating wiggle.
• Residual — whatever’s left after removing trend and seasonality; should look like noise.
• Additive vs. multiplicative — whether the seasonal swing is a constant amount or a constant percentage of the level.

Why This Matters#

Every technique later in this course either builds one of these three components or reacts to one. Differencing (Module 3) is a way of removing trend. SARIMA’s seasonal order (Module 6) is a formal model of the seasonal component. And the residual is, in a very real sense, the thing every forecasting model is ultimately trying to shrink to pure noise — the smaller and less structured it gets, the less is left to explain. Learning to name and separate these three pieces now is what makes every later diagnostic (“does this need differencing?”, “is the seasonal order right?”, “did the model miss something?”) a question you can actually answer by looking at a plot.

Next Steps#

Continue Building Your Skills#

You now have precise definitions for trend, seasonality, and residual, and you know the two ways they combine. Next you’ll build a decomposition from scratch on the real Cyclepath series — a centered moving average for the trend, subtraction to expose the season, and whatever’s left as the residual — so you see exactly what a decomposition function does before you ever call one.