Time series analysis sits at the centre of applied macroeconomics, finance, and energy economics. Whenever a single entity is observed over time — a country's GDP, an exchange rate, a commodity price, a firm's returns — the time dimension brings its own set of problems: trends, unit roots, autocorrelation, seasonality, structural breaks, and changing volatility. Stata handles all of it, but the commands are spread across a wide methodological range, from simple regression to ARDL, causality testing, error correction, and GARCH.
This playlist organises the workflow into a sensible learning path. Work through the stages in order. Each one assumes the previous: you cannot interpret an ARDL bounds test without first getting the series imported and stationarity sorted, and the logic of error correction only makes sense once the regression and cointegration foundations are in place.
Stages 1–4 form the core applied workflow, from data import through volatility and projection. Stages 5 and 6 are advanced extensions layered on that foundation — Fourier and quantile cointegration, mixed-order systems, and a specialised toolkit of quantile-on-quantile, wavelet, and visualisation methods — for researchers pushing beyond the standard path.
Foundations & Stationarity
Every time series analysis is only as reliable as the series behind it and the stationarity checks applied to it. Before any model runs, the data must be brought in cleanly, declared to Stata as time series, transformed appropriately, and tested for unit roots. This stage covers the unglamorous but decisive groundwork — getting data out of Excel, deciding between levels and logs, repairing gaps, and establishing the integration order of each series with modern, break- and quantile-robust unit-root tests.
Start here. A quick, practical guide for beginners on bringing time series data from Excel into Stata correctly — getting the date variable recognised, declaring the series with tsset, and avoiding the import mistakes that quietly break later commands.
Whether to model a series in raw levels or in logs is not a matter of taste — it changes interpretation and can change your results. This video explains how to use the kmtest command to decide formally whether a log transformation is appropriate for your data.
Gaps inside a series break lag operators and shrink your estimation sample. This video covers how to handle internal missing observations by interpolating with ipolate, restoring a continuous series ready for dynamic modelling.
Standard ADF and KPSS tests assume a fixed deterministic structure and lose power when a series carries smooth breaks, seasonality, or cyclical movement. This video applies Fourier-based unit-root tests through the fourierur command, which approximates unknown structural shifts with a small number of trigonometric terms — giving more reliable stationarity decisions for periodic, cyclic, and high-frequency data.
A series can be stationary in its centre but persistent in its tails, or the reverse — a single conditional-mean test hides this. This video runs the quantile unit-root test (Koenker–Xiao type), checking for a unit root at different positions of the conditional distribution rather than only at the mean. Useful when adjustment speed differs across calm and turbulent regimes.
Watch VideoRegression & ARDL Models
With clean, stationarity-checked data in hand, this stage builds the core modelling toolkit. It starts with simple time series regression and the diagnostics that must accompany it, then moves through the ARDL family — the workhorse for series mixing I(0) and I(1) variables — into its asymmetric, threshold, and quantile variants for nonlinear and distribution-dependent effects.
A comprehensive tutorial on running a basic time series regression and, crucially, performing the diagnostic tests that come with it — autocorrelation, heteroskedasticity, and stability checks. The diagnostics are not optional housekeeping; they determine whether the regression can be trusted at all.
Watch VideoA wide overview of the ARDL variants available in Stata, from the basic Pesaran bounds-test model to bootstrap, Fourier, asymmetric (NARDL), quantile, and hybrid combinations. Covers the Basic, Bootstrap, Fourier, Fourier-Bootstrap, Asymmetric (NARDL), Bootstrap-NARDL, Fourier-NARDL, Fourier-Bootstrap-NARDL, Quantile, and Fourier-Quantile ARDL specifications — a map of which variant fits which data problem.
Watch VideoMany relationships are asymmetric — a price rise and a price fall of equal size do not produce mirror-image responses. This video focuses on the NARDL model, which decomposes a regressor into its positive and negative partial sums to capture these asymmetric short- and long-run effects.
Watch VideoHigh-frequency series often carry periodic and cyclical structure that a standard model cannot absorb. This video teaches how to handle seasonality and cycles using the Fourier Quantile ARDL approach, combining Fourier terms for smooth structural shifts with quantile estimation across the conditional distribution.
Watch VideoAsymmetry is sometimes regime-dependent: the response changes not just with the sign of a shock but once it crosses a threshold — and there may be more than one. This video estimates a multiple-threshold nonlinear ARDL with the tnardll command, partitioning a regressor at several breakpoints to recover distinct short- and long-run effects within each regime.
This video combines two ideas already met in the course — quantile estimation and the positive/negative decomposition of NARDL — into a quantile nonlinear ARDL, run with the qardl command. It traces asymmetric short- and long-run effects across the entire conditional distribution of the dependent variable, revealing where in the distribution the asymmetry actually bites rather than only at the mean.
Causality & Dynamics
Estimating a relationship is not the same as understanding its direction or its dynamics over time. This stage covers the tools that go beyond a single fitted equation — formal causality testing robust to integration order, rolling-window estimation that reveals how relationships evolve, the error correction framework that ties short-run movements to a long-run equilibrium, and asymmetric causality that separates the influence of positive and negative shocks.
Granger causality testing is sensitive to the integration and cointegration properties of the data. This video demonstrates how to run several causality tests, including the robust Toda-Yamamoto approach that sidesteps pre-testing pitfalls, along with Fourier and quantile variants for breaks and distributional causality.
Watch VideoA single full-sample estimate hides the fact that relationships and stationarity shift over time. This video shows how to use rolling windows to track how unit-root behaviour, long-run coefficients, and causality direction change across sub-periods — often revealing instability a static model would miss entirely.
Watch VideoWhen variables are cointegrated, the error correction model separates the long-run equilibrium from the short-run adjustment back towards it. This step-by-step walkthrough of the two-step ECM is aimed at beginners. For the underlying theory, the accompanying clip covers the basic concepts of the ECM equations before you estimate.
Watch Video ECM ConceptsConventional Granger causality treats positive and negative shocks symmetrically, but cause and effect are often direction-dependent — bad news may transmit where good news does not. This video covers the family of asymmetric causality tests (Hatemi-J style), separating the causal impact of positive and negative cumulative shocks between series.
Watch VideoVolatility & Projection
This stage moves from modelling the mean of a series to modelling its variance, and then to communicating results. Financial and high-frequency data routinely show volatility clustering that the mean models in earlier stages assume away. It closes the core path with conditional volatility modelling and a tool for turning estimated relationships into clear, projectable visuals.
When volatility clusters — large changes followed by large changes — the constant-variance assumption fails. This video explains the concepts behind ARCH and GARCH models and walks through estimating them in Stata to capture time-varying conditional volatility in financial and high-frequency series.
Watch VideoA coefficient table rarely communicates a result as clearly as a well-built projection plot. This video introduces the regproject command for visualising and projecting regression results across cross-sectional, panel, and time series settings — turning estimates into figures ready for papers and presentations.
Advanced Cointegration
The standard ARDL and Engle-Granger machinery assumes a fairly clean world: no structural breaks, a single conditional mean, and variables that are at most I(1). Real macro and financial systems frequently violate all three. This stage gathers the advanced cointegration tools for those harder cases — smooth breaks via Fourier terms, distribution-dependent equilibria via quantile cointegration, and systems mixing I(0), I(1), and I(2) variables.
Long-run relationships can drift through gradual structural change that standard cointegration tests misread as no cointegration. This video pairs Fourier-based cointegration testing — using trigonometric terms to absorb smooth breaks — with the multiple-threshold NARDL, giving an advanced framework for detecting and modelling regime-dependent long-run equilibria.
Watch VideoCointegration need not hold uniformly across the conditional distribution. This video surveys a family of eight quantile cointegration tests, examining whether a long-run relationship exists — and how strong it is — at different quantiles, which matters when equilibrium behaviour differs between calm and turbulent periods.
Watch VideoStandard VARs assume a common order of integration and stumble when a system mixes I(0), I(1), and I(2) variables. This video presents a residual-based Fully Modified VAR that corrects for endogeneity and serial correlation, allowing consistent estimation and inference across variables of different integration orders.
Watch VideoSome macro and monetary series — price levels, nominal money — behave as I(2), and treating them as I(1) invalidates the usual tests. This video runs cointegration tests designed for systems with a mixed I(1) and I(2) structure, identifying the long-run relations that standard I(1) cointegration analysis would miss or distort.
Watch VideoSpecialised Methods & Visualisation
The final stage collects powerful methods that sit outside the linear-model mainstream, together with the visualisation tools that make complex results legible. Quantile-on-quantile regression and wavelet analysis expose structure that mean-based models cannot see; nomographs and a disciplined model-selection workflow turn that structure into something a reader — or a reviewer — can actually use.
Quantile-on-quantile regression generalises quantile regression by letting each quantile of the dependent variable respond differently to each quantile of the regressor. This video estimates the full quantile-on-quantile surface in Stata and renders it as a 3D plot, exposing relationships that are invisible to both OLS and ordinary quantile regression.
Watch VideoTime-domain and frequency-domain views each tell only half the story. Wavelet analysis decomposes a series — or the co-movement between two series — across both time and frequency at once, revealing how short- and long-run relationships strengthen or break down over specific periods. This tutorial walks through wavelet decomposition and coherence in Stata.
Watch VideoA nomograph turns a fitted model into a graphical calculator: predicted outcomes can be read straight off aligned scales without computing anything by hand. This video shows how to build nomograph / nomolog plots for least-squares regression models, a feature included in the regproject package — extending the projection toolkit introduced in Stage 4.
With so many time series specifications available, choosing among them should be systematic rather than habitual. This video lays out a disciplined process for selecting models using statistical indicators — information criteria, fit, parsimony, and diagnostic adequacy — so the final model is defensible to a reviewer rather than merely convenient.
Watch VideoA note on sequencing
Stage 1 is a prerequisite, not a suggestion. If your series is not imported and declared correctly with tsset, the lag and difference operators that every later command relies on will either fail or silently mislead. If you have not settled the levels-versus-logs and stationarity questions, an ARDL bounds test in Stage 2 cannot be interpreted.
Stages 3 and 4 carry more flexibility depending on your research focus. If your interest is the direction of influence between variables, the causality and ECM videos are the core. If you work with financial or high-frequency data, the ARCH–GARCH material is where the real questions live. Stages 5 and 6 are advanced extensions — reach for Fourier or quantile cointegration, mixed-order systems, quantile-on-quantile regression, or wavelets only once the foundation beneath them is solid. The stronger that foundation before specialising, the better your instincts will be when something goes wrong — and in time series, the trouble usually hides in the data preparation you were tempted to rush.
These methods are genuinely hard to use well. The places where they break — a missed unit root, an ignored break, an autocorrelated residual — are rarely obvious until a reviewer asks the question you should have asked yourself.