Does ACD require equally spaced observations?

No. It is designed around durations between irregular events.

Is psi the next actual duration?

No. It is the conditional mean in the stated model.

Does a shorter duration indicate faster activity?

Yes, for the same event definition and time units.

AI Glossary / Statistics and probability / Advanced

Autoregressive conditional duration

Definition

Autoregressive conditional duration, or ACD, is a time-series modeling framework for the positive waiting times between successive events. It lets the expected next duration depend on previous observed durations and previous conditional expectations. Introduced by Engle and Russell, it is especially associated with irregularly spaced transaction data, where activity can alternate between periods of short and long waits.

What it means

Many time series record values at fixed intervals, such as one observation each hour. Event data instead arrive whenever something happens. ACD models the elapsed time between those events rather than forcing them into equally spaced bins.

The event index advances one step per event, even though the corresponding amount of clock time varies. A period with short durations represents more frequent activity; a period with long durations represents less frequent activity.

The model predicts a conditional distribution of waiting times, not the exact arrival time of the next event. Its name refers to autoregression in duration dynamics, and should not be confused with a model directly predicting transaction prices or with ARCH models for conditional variance.

How it works

Let x_i be the duration between events i−1 and i. A basic specification writes x_i = psi_i epsilon_i, where psi_i is the expected duration given past information and epsilon_i is a positive innovation with conditional mean one.

In ACD(1,1), psi_i = omega + alpha x_(i−1) + beta psi_(i−1). With a positive omega, nonnegative alpha and beta, and positive initialization, the predicted duration stays positive. Alpha controls the response to the latest observed duration, while beta carries forward the previous conditional expectation.

A distribution for the positive innovations, such as a unit-mean exponential distribution, completes a particular conditional model. Parameters can be estimated by likelihood or appropriate quasi-likelihood methods. The innovation distribution determines more than the mean, including tail probabilities and the waiting-time hazard.

For the standard finite-mean stationary specification with alpha + beta below one, the unconditional mean duration is omega/(1−alpha−beta). Strict stationarity, finite moments, and estimator properties are distinct requirements; stationarity alone should not be used to assert every large-sample guarantee. Analysis also needs to address predictable intraday patterns, timestamp resolution, interruptions, and the fact that a fixed observation window contains a random number of events.

Three practical examples

A transaction-data researcher models clustering in the waits between trades without converting them into fixed-minute counts.
An operations analyst explores whether the intervals between logged service arrivals exhibit persistent short-wait and long-wait periods.
A simulation team uses fitted duration dynamics to generate event times for evaluating a system under varying activity levels.

Think of it this way

Think of waiting for messages in a busy group conversation. A recent burst of rapid replies may suggest another short wait, while a quiet stretch suggests a longer one. ACD formalizes this dependence in the waiting times, while still allowing the next gap to differ unpredictably from its forecast.

Worked example

Consider an original ACD(1,1) model measured in seconds. Let omega = 1, alpha = 0.2, and beta = 0.5. The previous observed duration is 8 seconds, and the previous conditional mean duration is 6 seconds.

The next conditional mean is 1 + 0.2 × 8 + 0.5 × 6 = 5.6 seconds. If the next positive innovation equals 1.5, the realized duration is 5.6 × 1.5 = 8.4 seconds. The model then updates the following conditional mean to 1 + 0.2 × 8.4 + 0.5 × 5.6 = 5.48 seconds.

For this finite-mean stationary specification, the long-run mean duration is 1/(1−0.2−0.5), approximately 3.333 seconds. It differs from the current 5.6-second conditional forecast because the current history reflects a relatively quiet period.

If the innovations are specifically unit-mean exponential, the probability that the next duration exceeds 10 seconds is exp(−10/5.6), approximately 0.168. This last probability depends on the exponential choice; the conditional-mean recursion alone does not determine it. All parameters here are stipulated, and the calculation is not an estimate from transaction data.

Illustrative case study

Fictional case: Seabrook Services analyzes timestamped support arrivals. It notices that short intervals tend to follow other short intervals, but also finds regular differences between morning and evening activity.

The team separates predictable time-of-day effects from remaining duration dependence before comparing an ACD model with simpler alternatives. It checks standardized durations and evaluates forecasts on later observation windows. Duplicate timestamps and overnight closures are handled under documented rules.

The fictional analysis informs a simulation of arrival timing. It does not establish that all arrivals follow an ACD process, and it does not translate a waiting-time forecast directly into an exact staffing requirement without considering service duration and workload variability.

Common misconceptions and mistakes

Common mistakes include treating event index as clock time, confusing conditional mean duration with a realized wait, and assuming every ACD model has exponential innovations.

Another error is applying the finite-mean formula when its conditions do not hold. A reciprocal expected duration also should not automatically be interpreted as an arbitrary model's instantaneous hazard; the hazard depends on the full conditional distribution and elapsed waiting time. Ignoring seasonal patterns, rounded timestamps, or observation-window boundaries can make apparent duration dependence misleading.

Business use

ACD models can help businesses understand and simulate irregular event activity. They may be useful when the timing of events affects system load, execution processes, or operational responsiveness.

Define the event precisely and choose a time unit and observation window appropriate to the decision. Compare conditional distributions, not only average waits, when extreme gaps or bursts matter. Validate the model against realistic alternatives and maintain a clear distinction between describing timing patterns and explaining the mechanisms that cause them.

Check your understanding

Why is the realized 8.4-second duration different from the 5.6-second forecast?

Answer: The realized duration multiplies the conditional mean by the positive innovation, which need not equal its mean of one.

Related terms

Duration
Conditional expectation
event time
point process
positive innovation
conditional intensity
quasi-maximum likelihood
GARCH

Learn next

compare duration and intensity models, inspect residual duration distributions, and study the effects of seasonality and random observation counts on inference.

Duration

Sources

Sources: Cavaliere, Mikosch, Rahbek, and Vilandt, Asymptotics for the Generalized Autoregressive Conditional Duration Model (2023), https://arxiv.org/pdf/2307.01779, for the ACD(1,1) formulation and finite-mean inference distinction; Engle and Russell's original model publication record, https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6809. The recursion, exponential tail calculation, and service case are original teaching illustrations.

Previous: Autorater evaluation Next: Autoregressive Conditional Heteroskedasticity