Skip to content

Autoregressive Conditional Duration

Model positive intervals between irregular events as a unit-mean innovation times a conditional expected duration that evolves from past durations and past conditional means.

Version
v2 · 2026-09-06 · History
Domain-specific #
1324
Origin domain
economics
Subdomain
financial econometrics
Aliases
ACD model, Autoregressive conditional duration model

Core Idea

Let event times satisfy \(t_i>t_{i-1}\) and define duration \(x_i=t_i-t_{i-1}>0\). An autoregressive conditional duration (ACD) model writes

\[ x_i=\psi_i\varepsilon_i,\qquad E(\varepsilon_i\mid\mathcal F_{i-1})=1, \]

where \(\psi_i=E(x_i\mid\mathcal F_{i-1})\) evolves from prior durations and conditional means. In a linear ACD\((p,q)\),

\[ \psi_i=\omega+\sum_{j=1}^{q}\alpha_jx_{i-j}+ \sum_{k=1}^{p}\beta_k\psi_{i-k}, \]

with positivity restrictions. Engle and Russell introduced the model for clustered, irregularly spaced financial transaction data.[1]

Structural Signature

  • An ordered sequence of event times.
  • Strictly positive inter-event durations.
  • An information set from past events and covariates.
  • A conditional expected duration \(\psi_i\).
  • A multiplicative positive innovation with unit conditional mean.
  • Autoregressive terms in realized durations.
  • Persistence terms in past conditional means.
  • Positivity and stationarity restrictions.
  • A declared innovation distribution and hazard shape.
  • Removal or modeling of deterministic intraday seasonality.
  • Likelihood estimation and standardized-residual diagnostics.
  • Forecasts of duration, intensity, or event probability.

What It Is Not

It is not an ordinary autoregression on equally spaced observations. It is not a renewal process unless the conditional mean is constant and durations become iid. It is not GARCH: the recursion is analogous, but the modeled object is conditional duration rather than conditional variance. It is not a homogeneous Poisson process, whose exponential durations lack autoregressive clustering.

Scope of Application

ACD models handle trade, quote, price-change, volume, and order-arrival durations in ultra-high-frequency finance. They also provide a template for irregular event sequences in other domains. Engle embeds duration analysis in marked point-process views of complete transaction records and links surprising duration to market-microstructure quantities.[2]

Clarity

Define the event, timestamp resolution, zero-duration treatment, overnight gaps, censoring, time-zone/session boundaries, and deterministic seasonality adjustment. State lag orders, recursion form, innovation distribution, covariates, positivity constraints, likelihood, and residual tests. Do not interpret a shorter predicted duration as higher activity without declaring the reciprocal intensity convention.

Manages Complexity

ACD models move analysis from arbitrary calendar bins to event time. A compact recursion summarizes duration clustering, while the innovation distribution separates predictable pace from unpredictable waiting. The same framework supports likelihood, hazard, forecast, and marked-event extensions.

Abstract Reasoning

  1. Define and clean the event sequence.
  2. Compute positive durations and handle ties.
  3. Estimate or model intraday seasonality.
  4. Select ACD lag orders and recursion family.
  5. Choose a positive unit-mean innovation distribution.
  6. Estimate parameters under admissibility constraints.
  7. Standardize durations by fitted \(\psi_i\).
  8. Test residual dependence and distributional fit.
  9. Compare forecasts and alternative duration models.

Knowledge Transfer

The portable pattern is factor a positive waiting time into predictable local pace and unit-scale surprise, then let pace inherit memory from earlier waits. It transfers to irregular event-stream forecasting. The proposed immediate parent is Stochastic Process.

Examples

In an ACD(1,1), a long observed trade interval raises the next conditional duration through \(\alpha x_{i-1}\), while \(\beta\psi_{i-1}\) carries persistence. Under the usual linear specification, \(\alpha+\beta<1\) supports a finite unconditional mean; values near one imply persistent duration clustering.

Exponential innovations impose a constant baseline hazard after scaling, while Weibull or generalized-gamma choices permit other shapes. Bauwens and Giot develop log-ACD and market-microstructure variants that relax positivity restrictions on linear predictors.[3] Surveys catalog nonlinear, semiparametric, multivariate, and marked extensions.[4]

Structural Tensions

  • Event time versus calendar time.
  • Flexible duration dependence versus parameter interpretability.
  • Linear positivity constraints versus log specifications.
  • Intraday seasonality versus stochastic persistence.
  • Hazard flexibility versus estimation stability.
  • Timestamp discreteness and zero durations versus continuous models.

Structural–Framed Character

Conditional factorization, autoregression, innovation scaling, persistence, and residual diagnostics are structural. Transaction durations, event time, intraday seasonality, hazards, and market microstructure are constitutive. The identity is domain-specific.

Structural Core vs. Domain Accent

The portable core is positive interval = predictable conditional scale × unit innovation. The domain accent is irregular high-frequency financial events and their clustered durations.

Stochastic Process is the proposed immediate parent. Stationarity, Autoregression, Renewal Process, Point Process, Time, and Forecasting are related. Renewal Process becomes a limiting boundary when duration dependence disappears.

The prospective queue contains one strict edge to prime:stochastic_process. No live DAG mutation is authorized.

Relationships to Other Abstractions

Local relationship map for Autoregressive Conditional DurationParents appear above the current abstraction, mutual partners to the right, and children below. Node labels state whether each abstraction is prime or domain-specific; colors identify relation types.AutoregressiveConditional DurationDOMAINPrime abstraction: Stochastic Process — is a kind ofStochasticProcessPRIME

Current abstraction Autoregressive Conditional Duration Domain-specific

Parents (1) — more general patterns this builds on

  • Autoregressive Conditional Duration is a kind of Stochastic Process Prime

    Stochastic Process is the proposed immediate parent.

Hierarchy path (1) — routes to 1 parentless root

Neighborhood in Abstraction Space

Autoregressive Conditional Duration sits in a sparse region of the domain-specific corpus (97th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (1565 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-09-08

Not to Be Confused With

  • GARCH conditional variance.
  • Equally spaced AR or ARMA model.
  • Renewal or homogeneous Poisson process.
  • Survival regression without event-time autoregression.
  • Deterministic intraday seasonality.
  • Raw trading intensity without a reciprocal convention.

References

[1] Robert F. Engle and Jeffrey R. Russell, “Autoregressive Conditional Duration: A New Model for Irregularly Spaced Transaction Data,” Econometrica 66 (1998): 1127–1162, doi:10.2307/2999632. registry ↩

[2] Robert F. Engle, “The Econometrics of Ultra-High-Frequency Data,” Econometrica 68 (2000): 1–22, doi:10.1111/1468-0262.00091. registry ↩

[3] Luc Bauwens and Pierre Giot, Econometric Modelling of Stock Market Intraday Activity (Kluwer, 2001), doi:10.1007/978-1-4757-3381-5. registry ↩

[4] Maria Pacurar, “Autoregressive Conditional Duration Models in Finance: A Survey,” Journal of Economic Surveys 22 (2008): 711–751, doi:10.1111/j.1467-6419.2007.00547.x. registry ↩