Aller au contenu principal

Cookies sur Macro by Mark.Nous utilisons des cookies essentiels pour la connexion, la langue et la sécurité. Avec votre accord, nous ajoutons des cookies optionnels de préférences, d’analyse et de mesure publicitaire avec le pixel Meta (Facebook). Lisez notre politique relative aux cookies.

Macro by MarkMacro by Mark
  • Accueil
Chargement…
Chargement…
← Modèles
MacroConceptsPolitiqueModèles

ARDL / bounds testing Model

Autoregressive distributed lag model for short-run dynamics and long-run relationships among variables that may have mixed integration orders.

Vue d’ensembleGraphiqueMathématiquesRéférences

Understand the mathematics

Follow the same equations and assumptions one section at a time or in a continuous reading.

Hover or focus a symbol defined in this model to find its exact matches. Select it to keep the highlight; Escape clears the selection.

Ardl Specification and Error-Correction Reparameterization

The ARDL(p,q1,…,qk)\text{ARDL}(p, q_1, \ldots, q_k)ARDL(p,q1​,…,qk​) model in levels is a standard dynamic regression with multiple lag orders. The key result is that it can be reparameterized into a conditional error-correction model (ECM) where the long-run coefficients appear explicitly as ratios of lagged-level coefficients. This reparameterization is algebraically exact, not an approximation.

We work with a single dependent variable yty_tyt​ and kkk regressors collected in xt=(x1t,…,xkt)′x_t = (x_{1t}, \ldots, x_{kt})'xt​=(x1t​,…,xkt​)′. The levels ARDL includes ppp lags of yyy and qjq_jqj​ lags of each xjx_jxj​.

yt=a0+a1t+∑i=1pϕiyt−i+∑j=1k∑l=0qjδjlxj,t−l+uty_t = a_0 + a_1 t + \sum_{i=1}^p \phi_i y_{t-i} + \sum_{j=1}^k \sum_{l=0}^{q_j} \delta_{jl} x_{j,t-l} + u_tyt​=a0​+a1​t+i=1∑p​ϕi​yt−i​+j=1∑k​l=0∑qj​​δjl​xj,t−l​+ut​

The standard ARDL in levels. Each regressor xjx_jxj​ enters with its own lag order qjq_jqj​.

Δyt=a0+a1t+πyyt−1+∑j=1kπjxj,t−1+∑i=1p−1ψiΔyt−i+∑j=1k∑l=0qj−1φjlΔxj,t−l+ut\Delta y_t = a_0 + a_1 t + \pi_y y_{t-1} + \sum_{j=1}^k \pi_j x_{j,t-1} + \sum_{i=1}^{p-1} \psi_i \Delta y_{t-i} + \sum_{j=1}^k \sum_{l=0}^{q_j - 1} \varphi_{jl} \Delta x_{j,t-l} + u_tΔyt​=a0​+a1​t+πy​yt−1​+j=1∑k​πj​xj,t−1​+i=1∑p−1​ψi​Δyt−i​+j=1∑k​l=0∑qj​−1​φjl​Δxj,t−l​+ut​
πy\pi_yπy​πj\pi_jπj​

The error-correction reparameterization. The lagged levels carry the long-run information; the differenced terms capture short-run dynamics.

β^j=−π^jπ^y=∑l=0qjδ^jl1−∑i=1pϕ^i\hat{\beta}_j = -\frac{\hat{\pi}_j}{\hat{\pi}_y} = \frac{\sum_{l=0}^{q_j} \hat{\delta}_{jl}}{1 - \sum_{i=1}^p \hat{\phi}_i}β^​j​=−π^y​π^j​​=1−∑i=1p​ϕ^​i​∑l=0qj​​δ^jl​​

The long-run multiplier for xjx_jxj​: the total (cumulative) effect of a permanent unit change in xjx_jxj​ on yyy in the long run.

Δyt=a0+a1t+πy(yt−1−β0−∑j=1kβjxj,t−1)+short-run terms+ut\Delta y_t = a_0 + a_1 t + \pi_y \bigl(y_{t-1} - \beta_0 - \sum_{j=1}^k \beta_j x_{j,t-1}\bigr) + \text{short-run terms} + u_tΔyt​=a0​+a1​t+πy​(yt−1​−β0​−j=1∑k​βj​xj,t−1​)+short-run terms+ut​
πy\pi_yπy​

The ECM written with the error-correction term in parentheses. The term (yt−1−β′xt−1)(y_{t-1} - \beta'x_{t-1})(yt−1​−β′xt−1​) is the lagged deviation from long-run equilibrium.

Start with the levels ARDL(p,q1,…,qk)\text{ARDL}(p, q_1, \ldots, q_k)ARDL(p,q1​,…,qk​)

The starting point is the standard ARDL in levels.
yt=a0+∑i=1pϕiyt−i+∑j=1k∑l=0qjδjlxj,t−l+uty_t = a_0 + \sum_{i=1}^p \phi_i y_{t-i} + \sum_{j=1}^k \sum_{l=0}^{q_j} \delta_{jl} x_{j,t-l} + u_tyt​=a0​+i=1∑p​ϕi​yt−i​+j=1∑k​l=0∑qj​​δjl​xj,t−l​+ut​

Add and subtract yt−1y_{t-1}yt−1​ on the left side

Introduce Δyt\Delta y_tΔyt​ on the left by subtracting yt−1y_{t-1}yt−1​.
yt−yt−1=a0+(ϕ1−1)yt−1+∑i=2pϕiyt−i+∑j=1k∑l=0qjδjlxj,t−l+uty_t - y_{t-1} = a_0 + (\phi_1 - 1) y_{t-1} + \sum_{i=2}^p \phi_i y_{t-i} + \sum_{j=1}^k \sum_{l=0}^{q_j} \delta_{jl} x_{j,t-l} + u_tyt​−yt−1​=a0​+(ϕ1​−1)yt−1​+i=2∑p​ϕi​yt−i​+j=1∑k​l=0∑qj​​δjl​xj,t−l​+ut​

Recursively substitute lagged differences

Repeated addition/subtraction of lagged yyy and lagged xxx terms collects the levels into πy\pi_yπy​ and πj\pi_jπj​, and the first differences into ψi\psi_iψi​ and φjl\varphi_{jl}φjl​. This is algebraic identity, not estimation.
Δyt=a0+πyyt−1+∑j=1kπjxj,t−1+∑i=1p−1ψiΔyt−i+∑j=1k∑l=0qj−1φjlΔxj,t−l+ut\Delta y_t = a_0 + \pi_y y_{t-1} + \sum_{j=1}^k \pi_j x_{j,t-1} + \sum_{i=1}^{p-1} \psi_i \Delta y_{t-i} + \sum_{j=1}^k \sum_{l=0}^{q_j-1} \varphi_{jl} \Delta x_{j,t-l} + u_tΔyt​=a0​+πy​yt−1​+j=1∑k​πj​xj,t−1​+i=1∑p−1​ψi​Δyt−i​+j=1∑k​l=0∑qj​−1​φjl​Δxj,t−l​+ut​
πy\pi_yπy​πj\pi_jπj​

Define the error-correction speed

This is the sum of the AR lag polynomial evaluated at z=1z = 1z=1. Stationarity of yyy requires the polynomial to have roots outside the unit circle, which guarantees πy<0\pi_y < 0πy​<0.
πy=−(1−ϕ1−ϕ2−⋯−ϕp)\pi_y = -(1 - \phi_1 - \phi_2 - \cdots - \phi_p)πy​=−(1−ϕ1​−ϕ2​−⋯−ϕp​)
πy\pi_yπy​

Define the long-run regressors' coefficients

The sum of all lag coefficients for xjx_jxj​. This is the total multiplier before normalizing by the AR polynomial.
πj=δj0+δj1+⋯+δj,qj\pi_j = \delta_{j0} + \delta_{j1} + \cdots + \delta_{j,q_j}πj​=δj0​+δj1​+⋯+δj,qj​​
πj\pi_jπj​

Recover the long-run coefficients

The long-run equilibrium: setting all differences to zero and solving for yyy in terms of xxx gives y=β0+∑(βjxj)y = \beta_0 + \sum(\beta_j x_j)y=β0​+∑(βj​xj​). This is the cointegrating relationship.
βj=−πj/πy=δj0+δj1+⋯+δj,qj1−ϕ1−ϕ2−⋯−ϕp\beta_j = -\pi_j / \pi_y = \frac{\delta_{j0} + \delta_{j1} + \cdots + \delta_{j,q_j}}{1 - \phi_1 - \phi_2 - \cdots - \phi_p}βj​=−πj​/πy​=1−ϕ1​−ϕ2​−⋯−ϕp​δj0​+δj1​+⋯+δj,qj​​​
βj=−πj/πy\beta_j = -\pi_j / \pi_yβj​=−πj​/πy​

Bewley (1979) transform for inference

Direct inference on βj=−πj/πy\beta_j = -\pi_j / \pi_yβj​=−πj​/πy​ is complicated by the ratio structure. Bewley (1979) proposed an instrumental-variables reparameterization that produces βj\beta_jβj​ directly as a regression coefficient with a standard t-statistic. The instrument is the fitted value from the unrestricted ARDL.

Transition checks

Sampled dynamics
zt=G(zt−1;θ)+ηt,yt=Hzt+εtz_t=G(z_{t-1};\theta)+\eta_t,\qquad y_t=H z_t+\varepsilon_tzt​=G(zt−1​;θ)+ηt​,yt​=Hzt​+εt​

The state is updated at observed dates. A line between dates is a display convention unless the model specifies behavior inside the interval.

Read adjacent points as adjacent sample calculations. Smoothness is a property of the estimator, not of the chart line.

Do not infer within-period behavior from a sampled estimator.

Variable timing

Latent state

zt=G(zt−1;θ)+ηtz_t=G(z_{t-1};\theta)+\eta_tzt​=G(zt−1​;θ)+ηt​

The state updates at sample dates. The visible line is only as smooth as the estimated state sequence.

Observation equation

yt=Hzt+εty_t=Hz_t+\varepsilon_tyt​=Hzt​+εt​

Observed series can move differently from latent states because measurement noise and loadings matter.

Keep going

Read next

Continue with a related concept, model, or example.

Theory model

IS-LM Model

a well-connected theoretical model two steps from here

New angle: theoretical model →
Macro by Mark
Aide
Politique de confidentialitéConditions d’utilisationÉthique et conformité

© 2026 Mark Jayson Nation LLC