Three Versions of Time Series

A decision guide to classical statistical, financial, and economic time series, with matched examples and method-selection rules.

Three Versions of Time Series

The shortest useful distinction

All three fields study ordered random variables. They differ in what makes an analysis credible:

Question	Classical statistical time series	Financial time series	Economic time series
primary object	stochastic process and dependence structure	returns, volatility, liquidity, and tail loss	growth, cycles, long-run relations, and policy transmission
common clock	regular, fixed interval	trading time; daily to tick-by-tick	release time; monthly/quarterly, mixed frequency
usual transformation	centre, detrend, seasonally adjust	adjusted price $\rightarrow$ log return	level $\rightarrow$ log, difference, growth, gap
dominant difficulty	valid covariance and stable dynamics	weak mean signal, changing variance, heavy tails	persistence, unit roots, revisions, latent current state
validation target	model adequacy and prediction error	economic value and tail calibration after costs	real-time forecast or credible structural interpretation
costly error	invalid stochastic model	underestimated loss or false predictability	bad nowcast or misidentified policy effect

The classical course asks, “What follows from this stochastic model?”

The applied course asks, “Is this model aligned with the data available and the decision being made?”

What remains unchanged

For the finite vector $\mathbf y=(y_1,\ldots,y_T)^\top$ , the common language is still:

\mathbf y=X\beta+\mathbf u,\qquad E(\mathbf u\mid X)=0,\qquad \operatorname{Var}(\mathbf u\mid X)=\Omega.

Classical analysis studies what structures such as Toeplitz $\Omega$ imply.
Finance often allows diagonal elements of $\Omega$ to evolve through time and evaluates tail functionals.
Economics often models persistent means, common trends, simultaneous systems, and imperfectly observed states.

The algebra transfers. The information set, interpretation, and loss do not transfer automatically.

One method, three uses

Method	Classical use	Finance use	Economics use
AR/ARMA	represent stationary dependence	short-horizon return or spread dynamics	inflation, growth, or forecast benchmark
HAC covariance	inference with serial correlation	overlapping multi-period returns	distributed lags and persistent macro regressors
GARCH	example of nonlinear conditional variance	volatility, VaR, derivative/risk inputs	inflation or exchange-rate uncertainty when relevant
cointegration	reduced-rank long-run system	spreads, term structure, price discovery	money–prices, consumption–income, output relations
VAR	multivariate forecasting	return–volatility–liquidity interactions	policy transmission and macro forecasting
Kalman filter	efficient state recursion	latent volatility, beta, or efficient price	nowcasting, output gaps, mixed-frequency factors
spectral methods	frequency decomposition	cycles in volatility or market activity	business-cycle and seasonal frequency separation

A VAR can forecast without identifying a structural shock. A stationary spread can exist without yielding a profitable trade. A significant return predictor can fail after transaction costs. Never upgrade a statistical relation into an economic claim without the missing assumptions.

Matched example 1: a persistent level

Suppose

x_t=x_{t-1}+\eta_t.

In the classical course, this is a unit root: shocks have permanent effects and the level is nonstationary.

In finance, $x_t$ may be a log price. We usually analyse the return

r_t=\Delta x_t=\eta_t,

because it is the investable one-period gain and is closer to stationary.

In economics, $x_t$ may be log real GDP. Differencing produces growth, but discards the level relation needed for questions about potential output or cointegration. The transformation must follow the estimand, not a stationarity ritual.

Matched example 2: a VAR innovation

Write a reduced-form VAR(1):

\mathbf y_t=A\mathbf y_{t-1}+\mathbf u_t, \qquad E(\mathbf u_t\mathbf u_t^\top)=\Sigma_u.

The forecast response to $\mathbf u_t$ is well defined. A named shock requires

\mathbf u_t=B\boldsymbol\varepsilon_t,\qquad E(\boldsymbol\varepsilon_t\boldsymbol\varepsilon_t^\top)=I,\qquad BB^\top=\Sigma_u.

There are many matrices $B$ satisfying the last equality.

A finance study may order returns before liquidity and interpret a recursive response cautiously.
A macro study may use timing restrictions, sign restrictions, or an external instrument to identify a monetary-policy shock.
The reduced-form fit alone cannot decide which $B$ is economically correct.

Jordà's local projections estimate horizon-specific responses directly. Plagborg-Møller and Wolf show that unrestricted local projections and VARs target the same impulse responses; practical differences arise from regularisation, lag choices, and finite samples, not from automatic identification.

Matched example 3: the forecast loss

Let $e_{t+h}=y_{t+h}-\hat y_{t+h\mid t}$ .

A classical exercise may minimise $E(e_{t+h}^2)$ .
A portfolio desk may care more about underpredicting the lower tail than a symmetric mean error.
A policy institution may care about forecast revisions before a meeting, conditional performance during recessions, or a density covering multiple scenarios.

The correct model comparison is therefore

E[L(y_{t+h},\hat f_{t+h\mid t})],

where $L$ is chosen before seeing the winning model.

The usage decision

Use this sequence before fitting anything:

Decision	If yes	Consequence
Is the object a traded price?	transform to an adjusted return unless the level relation is itself the target	inspect market calendar, corporate actions, costs, and tails
Is the target a macroeconomic level?	test whether trends and long-run relations are substantively meaningful	compare difference, cointegration, and state-space specifications
Was the value revised after release?	preserve vintage and release timestamps	evaluate against information available at each forecast origin
Is the claim causal or structural?	forecasting fit is insufficient	state and defend an identification design
Is the loss asymmetric or tail-focused?	RMSE is insufficient	evaluate quantiles, VaR/ES, or decision-specific utility
Do horizons overlap?	residuals share observations	use an appropriate long-run covariance and honest split

Three common category errors

Stationary therefore useful: a spread may be stationary but too slow, costly, or unstable to trade.
Predictive therefore causal: a yield spread may forecast activity without representing an intervention.
Revised therefore known: a final macro series can make a historical nowcast look better than information available in real time.

Practice

For each case, choose the track and the missing safeguard.

Daily close-to-close equity returns are used to estimate tomorrow's 1% loss quantile.
Quarterly consumption and income levels are modelled jointly to study long-run adjustment.
Monthly inflation and policy rates are used to report the response to a monetary-policy shock.
An ARMA simulation is used to verify a Yule–Walker identity.

Answers

Finance: adjusted prices, tail loss, changing volatility, and out-of-sample VaR calibration.
Economics with cointegration: test rank and interpret the error-correction relation.
Economics with structural identification: define why the innovation is a policy shock.
Classical statistics: covariance validity and algebra are the main objects.

Working rule

Use the classical course to justify the stochastic machinery. Use this course to justify the transformation, information set, identification, and decision. A strong analysis needs both.

Next: Data, Clocks, and Transformations.

Edit this pageorReport an issue

Financial and Economic Time Series — Course Guide

A matrix-based course in returns, volatility, persistent predictors, cointegration, structural dynamics, nowcasting, and forecast evaluation.

Data, Clocks, and Transformations

Transform prices and macroeconomic releases with explicit differencing, aggregation, timing, and vintage operators.