What Makes Data a Time Series?

P7.ts-foundations.01 · Audience: guest, it-ml, language-pro

Training Dynamics & Regularisation Clinic Trend, Seasonality & Decomposition

Everything in Traditional Machine Learning rested on one quiet assumption: the rows of your table are exchangeable. Shuffle them, split them, sample them — nothing of value is lost, because each row is a self-contained observation. This pillar is what happens when that assumption is dropped. In a time series the rows come in a fixed order, and the order itself carries the information: yesterday genuinely helps predict today. Shuffle a time series and you have not tidied your data — you have destroyed the signal.

Step 1 / 5 — Order is the signal

A time series is a sequence of measurements of the same thing, taken at regular intervals, kept in order: daily sales of a shop, hourly temperature at a weather station, monthly passengers on an airline. Two properties separate it from the tabular data of the previous pillar. First, the observations are not independent — today's sales are strongly related to yesterday's. Second, the position of a value matters — "sales were 120" means something different in December than in a flat week of February.

🗣️ From a linguist's perspective: A time series is a text, not a bag of words

The tabular data of the previous pillar was like a bag of words: you could pour the rows out in any order and count what you found. A time series is like a sentence — the same words in a different order mean something else, or nothing at all. "The dog bit the man" and "the man bit the dog" contain identical words; only the order distinguishes them. In this pillar, order is never packaging to be discarded — it is the grammar of the data.

ⓘ Concept: Time series

y(1), y(2), …, y(t), … — one value per time step, in a fixed order

Why it matters — Almost every business quantity worth forecasting — sales, traffic, demand, temperature, exchange rates — arrives as a time series. Treating it as ordinary tabular data silently breaks both the model and, worse, the evaluation that was supposed to catch the broken model.

✅ Key takeaways

A time series is ordered measurements of one thing over time. Unlike the tabular data of P5, the rows are not interchangeable — the order itself carries the signal.
Shuffling a time series before a train/test split leaks the future into training: the model interpolates between neighbours instead of forecasting, and the test score becomes an honest-looking lie.
Every series decomposes into level, trend, seasonality, and noise. Forecasting means capturing the first three and refusing to chase the fourth.
Repetition makes seasonality visible through noise; short windows hide gentle trends.
General regression and metrics stay with P5 — this pillar owns what the time axis adds: temporal structure and time-aware evaluation.

Next: P7.ts-foundations.02 — Trend, Seasonality & Decomposition

Ask the mentor about this module

Ask a question about this content. The mentor explains and grounds its answer in what you are studying; asking is recorded as a learning signal, not a grade.

Ctrl/Cmd + Enter to send

Try it yourself

A scratch console for this page's ideas — ungraded, nothing you run here is recorded.

Scratch console

A scratch console with the scientific stack (pandas, numpy, scikit-learn). Runs on the server — no network, resource-limited and measured.

Output appears here.

Where next?

Continue

Trend, Seasonality & Decomposition →

Time-Series Foundations · P7

Later in Time-Series Foundations

This module unlocks

Baselines You Must Beat

Go up a level