Financial Econometrics & Point Processes Nobel Prize Foundation Engle & Russell (1998)

Engle-Russell (1998) Autoregressive Conditional Duration (ACD) Lab

Model irregularly spaced transaction data in financial markets. Explore autoregressive trade arrival timing, Weibull baseline hazard functions, duration clustering, and the structural link between trade intensity and instantaneous return volatility.

Theoretical Foundations: Point Processes, Duration Clustering & Hazard Dynamics
1. The ACD(1,1) Specification

Let $t_i$ denote the timestamp of the $i$-th transaction, and $x_i = t_i - t_{i-1}$ be the observed duration between trades. The conditional expected duration $\psi_i = \mathbb{E}[x_i \mid \mathcal{F}_{i-1}]$ evolves according to:

$$\psi_i = \omega + \alpha x_{i-1} + \beta \psi_{i-1}$$

The multiplicative error specification decomposes duration as $x_i = \psi_i \cdot \epsilon_i$, where $\epsilon_i$ are i.i.d. innovations with $\mathbb{E}[\epsilon_i] = 1$. Covariance stationarity requires $\alpha + \beta < 1$, yielding the unconditional mean duration $\bar{x} = \frac{\omega}{1 - \alpha - \beta}$.

2. Weibull Hazard Rate & Duration Dependence

Under Weibull ACD (WACD), the standardized innovations follow $\epsilon_i \sim \text{Weibull}(\gamma, \lambda_0)$ where $\lambda_0 = 1 / \Gamma(1 + 1/\gamma)$. The conditional hazard rate $h(t \mid \psi_i)$ is:

$$h(t \mid \psi_i) = \frac{\gamma}{\psi_i} \left(\frac{t}{\psi_i}\right)^{\gamma - 1} [\lambda_0]^\gamma$$

$\gamma > 1$ implies positive duration dependence (hazard increases over time); $\gamma < 1$ implies negative duration dependence (hazard decreases, producing bursty cascades); and $\gamma = 1$ collapses to the memoryless exponential Poisson process.

Institutional Market Regimes & Presets

High-Frequency Algorithmic Clustered Flow: Strong duration persistence (alpha+beta = 0.97) and positive duration dependence (gamma = 1.25). Rapid burst clusters.

ACD Model Parameters
Persistence ($\alpha+\beta$)
0.970
Stationary
Uncond. Duration ($\bar{x}$)
1.67 s
Long-Term Mean
Duration Half-Life
22.8 trades
Shock Memory
Hazard Regime
-142.3
Increasing Hazard
Residual Diagnostics & Specification Tests
Mean Residual $\mathbb{E}[e_i]$
1.002
Theoretical = 1.000
Variance Residual $\text{Var}(e_i)$
0.684
EACD Benchmark = 1.000
Ljung-Box $Q(10)$ Test
8.42
No Residual Clust. ($p > 0.05$)
Sensitivity 1: Shock Impact ($\alpha$) vs Persistence ($\beta$) on Half-Life
Sensitivity 2: Expected Duration ($\psi$) vs Shape ($\gamma$) on Hazard Rate $h(1\text{s})$
ACD Point Process & Market Microstructure Mastery Quiz
Related Quantitative Finance & Execution Simulators
Kyle-Obizhaeva (2016) Invariance Lab

Model Pete Kyle & Anna Obizhaeva's universal invariance laws, bet size scaling, and square-root market impact.
Hasbrouck (1991) Trade VAR Lab

Measure permanent and transitory price impact of trades and private information variance using Vector Autoregression.
VPIN & Order Flow Toxicity Lab

Model Easley-López de Prado-O'Hara volume clock toxicity, bulk volume classification, and flash crash cascades.
EKOP (1996) PIN Lab

Estimate the Probability of Informed Trading via daily Poisson event trees and maximum likelihood optimization.
Avellaneda-Stoikov Market Making Lab

Model high-frequency market making, reservation price skewing, optimal bid-ask spreads, and inventory risk control.
GARCH(1,1) Volatility Forecasting Lab

Model autoregressive conditional heteroskedasticity, volatility clustering, and long-term variance reversion.
Hawkes (1971) Self-Exciting Process Lab

Model self-exciting point processes, Ogata thinning, branching ratio criticality, and endogenous flash crash cascades.
Frequently Asked Questions: Autoregressive Conditional Duration (ACD)

The Autoregressive Conditional Duration (ACD) model, developed by Nobel laureate Robert F. Engle and Jeffrey R. Russell (1998, Econometrica), is a foundational statistical framework for modeling irregularly spaced financial transaction data. Unlike standard time series that aggregate data into fixed calendar intervals (e.g., 1-minute or daily bars), the ACD model directly treats the duration between consecutive trades as a stochastic point process with autoregressive memory.

In the benchmark ACD(1,1) specification, the expected duration until the next trade $\psi_i = \mathbb{E}[x_i \mid \mathcal{F}_{i-1}]$ is modeled as a linear autoregression: $\psi_i = \omega + \alpha x_{i-1} + \beta \psi_{i-1}$, where $x_{i-1}$ is the most recently observed duration and $\psi_{i-1}$ is the previous expected duration. The model is stationary when $\alpha + \beta < 1$, giving an unconditional expected duration of $\mu = \frac{\omega}{1 - \alpha - \beta}$.

The difference lies in the assumed distribution of standardized residuals $\epsilon_i = x_i / \psi_i$. Exponential ACD (EACD) assumes $\epsilon_i \sim \text{Exp}(1)$, which implies a memoryless flat hazard rate. Weibull ACD (WACD) assumes $\epsilon_i$ follows a Weibull distribution with shape parameter $\gamma$, allowing for non-constant hazard rates that capture positive or negative duration dependence.

When $\gamma = 1$, WACD reduces to EACD with a constant hazard. When $\gamma > 1$, the baseline hazard rate increases over time (positive duration dependence), meaning the probability of an imminent trade grows the longer the market has waited. When $\gamma < 1$, the hazard rate decreases over time (negative duration dependence), characterizing bursty markets where trades cluster tightly, interspersed with long quiet periods.

As established by Easley and O'Hara (1992) and formalized econometrically by Engle (2000), informed traders enter the market in rapid succession to capitalize on perishable private information before prices adjust. Consequently, short trade durations (high trade intensity $\lambda = 1 / \psi$) directly signal an influx of informed flow, inducing an immediate surge in instantaneous price volatility.