Time Series

Data ordered in time, and why the order changes everything.

Also known as: time series, temporal data, time-ordered data

A time series is data with an order that matters: daily revenue, hourly error rates, monthly active users. The order is the whole point — shuffling the rows of a time series destroys its meaning, which is what separates it from ordinary tables where rows are interchangeable. Almost every number a business watches is a time series.

ordinary table:  rows interchangeable → summaries, group-bys
time series:     order is meaning → trends, seasons, streaks, breaks

That order breaks tools built for independent rows. Random train-test splits leak the future into training; averages hide movement; a single summary number cannot show a break halfway through. Time-series work starts by plotting the series as a line chart and reading its shape before computing anything.

The classic mistakes:

Treating time-ordered rows as independent. Correlated neighbours, drift and breaks violate the assumptions behind ordinary averages and tests. Methods must respect the order (see stationarity).
Summarising before looking. A monthly average can hide a mid-month collapse and recovery. Plot first, summarise second.
Comparing adjacent periods naively. This week vs last week mixes trend and seasonality into one number. Compare like with like, or split the components.
Future leakage. Any feature, split or aggregate that uses information from after the decision point inflates backtests and fails live. Cut strictly by time.

When it matters most: monitoring, alerting, forecasting and any before/after claim. If time is on an axis, the order is load-bearing — analyse it as a series, not a bag of rows.