What Is a Time Series?
Most of statistics begins with a bag of measurements you could shake up without losing anything: the
heights of a hundred people, the yields of fifty fields. Shuffle them and every average, every
histogram, every test comes out the same. A time series is the opposite kind of
data — shuffle it and you destroy it. The order is the information.
Formally, a time series is a sequence of observations
x_1, x_2, \dots, x_T recorded in order at times
t = 1, 2, \dots, T. A closing price each trading day, the temperature each
hour, the number of births each month, the voltage across a nerve each millisecond. The single defining
feature is serial dependence: the value now is statistically tied to the values just
before it. That one fact is what breaks classical "independent observations" statistics and forces an
entire subject of its own.
Four questions the subject answers
Everything in this course is aimed at one of four practical goals — and they are genuinely different
goals, not four names for the same thing:
- Describe. Split the series into interpretable pieces — trend, season, cycle,
noise — and summarize its dependence with the autocorrelation function.
- Forecast. Predict future values x_{T+1}, x_{T+2}, \dots
from the past, with a quantified uncertainty, not just a point.
- Explain / control. Understand the mechanism generating the series, or intervene
to steer it (a thermostat, a central bank, a controller).
- Detect. Flag when the series has changed — an anomaly, a structural break, a new
regime.
A weather model forecasts; a fraud system detects; an economist explains;
an engineer controls. Different goals, but all rest on the same core idea: capture the
dependence between a value and its own past.
What time series look like
The three curves below are stylized versions of the shapes you meet again and again. The
rising line is a trending series (think atmospheric CO₂). The
wavy one is seasonal — the same pattern repeating on a fixed period (retail
sales peaking each December). The jagged one wanders with no trend and no season: a
random walk,
the archetype of an unpredictable path like a stock price.
Notice what they share: each point sits close to the one before it. That closeness — nearby times take
nearby values — is serial dependence made visible, and measuring it precisely is the whole game.
Discrete time, and the sampling interval
We work almost entirely in discrete time: the index t steps
through the integers, one tick per observation, whether the underlying process is truly discrete (daily
rainfall totals) or a continuous signal sampled at a fixed rate (a temperature probe read every
minute). The gap between observations — the sampling interval — is a modelling choice
with teeth: sample too coarsely and you miss fast structure; the fastest wiggle you can even represent
is set by how often you look, a fact spectral analysis will make exact.
A series can also be univariate (one number per time, like a single thermometer) or
multivariate (a vector per time, like temperature and pressure and
humidity together). Most of this course is univariate; the
final module
handles the vector case, where series also influence one another.
A tempting first instinct is to fit x_t = \beta_0 + \beta_1 t + \varepsilon_t
by ordinary least squares and call it a day. Sometimes that captures the trend — but the standard
errors it hands you are usually lies. OLS assumes the residuals are independent; in a time
series they are almost never independent — a positive error today makes a positive error tomorrow more
likely. That leftover autocorrelation means your effective sample size is far smaller than
T, so OLS reports confidence intervals that are far too narrow and
significance that isn't there. Time series methods exist precisely to model the dependence OLS pretends
away — not as a decoration, but because ignoring it makes you confidently wrong.
In the 1920s the statistician Udny Yule was puzzling
over the sunspot numbers — a famously regular-ish cycle of roughly eleven years. Everyone modelled such
cycles as a hidden sine wave buried in noise. Yule's radical idea: maybe there is no fixed hidden
sine at all. Maybe each year's value is just a weighted echo of the previous couple of years, plus a
fresh random shock — a pendulum being kicked at random. That single reframing became the
autoregressive
model, the cornerstone of the whole field. The sunspots taught us that a series can look
cyclic without any clock ticking underneath — the "cycle" can be manufactured entirely by memory and
chance.