Time Seriesintermediate

The Nature of Time Series Data

A time series is a sequence of observations on the same unit indexed by time, such as quarterly GDP or daily returns. Unlike a cross section, the ordering matters and the data are best viewed as one realization of a stochastic process: the path we observe is a single draw from the joint distribution that nature could have generated. Because today’s value often depends on its own past, the independence we assume across a random sample breaks down, and the cross-section toolkit needs adapting. The sample size is the number of time periods $n$ , and we typically have just one history to learn from.

Why it matters

Imagine rewinding the economy and letting it run again. You would get a different GDP path each time, yet we only ever see one. That single path is our sample, so we cannot average over many independent units the way we do with a cross section. The catch is temporal dependence: this quarter’s unemployment is tied to last quarter’s, so observations carry information about their neighbors instead of being independent draws.

Formulas

A stochastic process

\{y_t : t = 1, 2, \dots, n\}

The observed series is one realization of this sequence of random variables.

A static regression on time series

y_t = \beta_0 + \beta_1 x_t + u_t

Same form as cross-section OLS, but indexed by time t and subject to its own assumptions.

Worked examples

Scenario

You load 40 years of quarterly Australian GDP growth and inflation into Stata to study their relationship.

Solution

First declare the data as time series so lag operators and date logic work: `tsset date, quarterly`. Then a simple model is `regress gdpgrowth inflation`. Stata now knows the temporal ordering, which lets you write lagged terms like `L.inflation` and run unit-root tests later.

NoteWithout `tsset`, the L. and D. time-series operators are unavailable.

Common mistakes

✗Time series can be analyzed exactly like cross-section data. The ordering carries information, so reshuffling rows destroys the autocorrelation structure that drives most time-series methods.
✗More years of data always means more independent information. Because observations are serially dependent, the effective information grows more slowly than the raw count of periods suggests.
✗We observe many realizations of the process. In practice we usually see only one path, so consistency rests on the process settling into a stable pattern over time rather than on repeated independent samples.
✗The error term cannot be correlated over time. In time series, serial correlation in $u_t$ is common and is one of the main complications we have to handle.

Revision bullets

•A time series is one realization of a stochastic process
•Temporal ordering matters; rows are not exchangeable
•Observations are typically serially dependent, not independent
•Sample size $n$ is the number of time periods
•Declare data with `tsset` before using lag and difference operators

Quick check

Why is the random sampling assumption from cross-section analysis usually inappropriate for time series?

Describing a time series as "one realization of a stochastic process" means:

Connected topics

Data Types Static & FDL Trends TS.1-TS.6 Unit Roots

Sources

Wooldridge (2019), Ch. 10
Wooldridge, Jeffrey M. Introductory Econometrics: A Modern Approach. 7th ed. Cengage, 2019.
Chapter 10 introduces the nature of time series data and the stochastic-process interpretation of a single realization.
Stock & Watson (2019), Ch. 14
Stock, James H., and Mark W. Watson. Introduction to Econometrics. 4th ed. Pearson, 2019.
Parallel treatment of time series data structures and why temporal ordering changes the analysis.

How to cite this page

Dr. Phil's Quant Lab. (2026). The Nature of Time Series Data. Derivatives Atlas. https://phucnguyenvan.com/concept/efm-ts-data

← Back to the atlas See in the network →