Tutorial · Intermediate

Time Series Analysis

Explore temporal structure, stationarity, lag dynamics, model validation, and forecasting in economic data.

40 min4 sections

Overview

What you will learn

Explore temporal structure, stationarity, lag dynamics, model validation, and forecasting in economic data.

  • Recognize trend, seasonality, and structural change
  • Understand stationarity and differencing
  • Select and diagnose dynamic models
  • Produce and evaluate forecasts

Step 01

Plot the series first

Time ordering contains essential information, so begin with a line plot before calculating correlations or fitting models.

1

Look for trend, seasonal cycles, volatility changes, unusual observations, and possible breaks.

2

Confirm the observation frequency and identify missing or irregular dates.

3

Use domain knowledge to distinguish data problems from genuine economic events.

Step 02

Check stationarity

Many time-series methods assume that the statistical behavior of the series is stable through time.

1

Inspect the level series and its autocorrelation function.

2

Use unit-root tests together with plots and economic reasoning.

3

Apply transformations or differencing when appropriate while preserving interpretable relationships.

Keep in mind

Over-differencing can remove meaningful long-run information and introduce unnecessary noise.

Step 03

Model lag dynamics

Current economic outcomes often depend on their own history and on delayed effects from other variables.

1

Use ACF and PACF patterns as initial guidance for autoregressive and moving-average terms.

2

Compare plausible specifications with information criteria and theoretical restrictions.

3

Consider ARDL or error-correction models when short-run dynamics and long-run relationships both matter.

Step 04

Validate and forecast

A fitted time-series model is useful only when its remaining errors behave like unpredictable noise.

1

Inspect residual autocorrelation and use a portmanteau test such as Ljung–Box.

2

Evaluate forecasts with a chronological holdout sample rather than a random split.

3

Report prediction intervals and update the model as new observations arrive.