Clock-Advance ETAS

• Clock-advance ETAS is a reformulation of the classical ETAS model using Lévy processes, removing the need for an artificial lower-magnitude cutoff.
• It implements local time advances to trigger aftershocks, enhancing temporal rescaling and computational efficiency in high-resolution earthquake catalogs.
• The iterative framework and advanced optimization methods improve forecast accuracy and interpretability in spatiotemporal seismic modeling.

Clock-advance ETAS refers to a reformulation of the classical Epidemic-Type Aftershock Sequence (ETAS) model, where earthquake sequences are understood through local time advances of independent background point processes. In this framework, triggered (aftershock) events are represented as “clock advances” imparted to independent Lévy (pure-jump) processes, yielding an ETAS model that is well-defined for arbitrarily small magnitudes without the need for ad hoc lower cutoffs. This approach also provides a foundation for flexible temporal rescaling in spatiotemporal earthquake forecasting, improving interpretability, computational tractability, and predictive quality in applications involving micro-seismicity and high-resolution catalogs (Holschneider, 2 Jan 2025, Das et al., 30 May 2025).

1. Classical ETAS Formulation

The classical ETAS model describes the temporal (and potentially spatial) sequence of earthquakes as a conditional point process. The one-dimensional (temporal) intensity function is

$$\lambda(t\mid \mathcal H_t)=\mu+\sum_{t_i<t} K\,h(t-t_i),$$

where $\mu$ is the background rate, $K$ is the productivity coefficient, and $h(t)=(1+t/c)^{-p}$ is the Omori-type triggering kernel. In the marked (magnitude-tracked) version,

$$\lambda(t,m\mid \mathcal H_t)=g(m)+\sum_{t_i<t} h(t-t_i) F(m_i)g(m),$$

where $g(m)$ is the Gutenberg–Richter magnitude density and $F(m_i)=10^{\kappa+\alpha m_i}$ describes aftershock productivity per ancestor event.

The standard ETAS framework requires an explicit lower-magnitude cutoff $m_0$ to ensure integrability, particularly in the branching structure where each earthquake may trigger an expected infinite number of direct aftershocks if the productivity law allows unbounded contributions from very small events (Holschneider, 2 Jan 2025). This limitation is addressed by the clock-advance approach.

2. Clock-Advance Interpretation and Lévy Process Representation

In the clock-advance ETAS, background seismicity is modeled as a pure-jump Lévy (Poisson) process in productivity-moment space. Given the transformation $Y=F(m)=10^{\kappa+\alpha m}$, the corresponding jump-size density is

$\psi(Y) = \gamma Y^{-1-\frac{b}{\alpha}},$

with normalization ensuring finite mean productivity for $\alpha>b$.

Each triggered event is interpreted as locally advancing the clock of an otherwise independent background realization. Specifically, aftershocks from an event at time $\tau$ with size $Y(\tau)$ are drawn from a background copy read out on the clock $H(t-\tau)=\int_0^{t-\tau}h(s)ds$, with an amplitude scaling by $Y(\tau)$. This constructs first-generation triggered events as $Y^{T,1}(t)=Y^B(Y(\tau) H(t-\tau))$, where $Y^B$ is an independent background jump process (Holschneider, 2 Jan 2025).

When multiple ancestors exist, the clock for their combined aftershock process is $H^*(t)=\sum_\tau Y(\tau)H(t-\tau)$, equivalent to the convolution $(h*C_Y)(t)$, where $C_Y(t)$ is the background cumulative productivity function.

3. Iterated Clock-Advances and Convergence

The full ETAS catalog is expressed as an iterated sum of generations, each realized by a Markov chain of clock-advanced, independent Lévy processes: $$Y^{ETAS}(t) = Y^{T,0}(t) + \sum_{k=1}^\infty Y^{T,k}(t),$$ with cumulative productivity

$$C_{Y}^{ETAS}(t) = C_{Y}^{T,0}(t) + \sum_{k=1}^\infty C_{Y}^{T,k}(t).$$

Each generation involves a recursion: given $Y^{T,k}$, its associated clock is $H^{*,k}(t)=(h*C_{Y}^{T,k})(t)$, and the next generation is a fresh independent background realization read out on $H^{*,k}(t)$. The Markov chain state evolution is

$$\begin{cases} E_{k+1}(t) = E_k(t) + W_k(t), \ W_{k+1}(t) = Y^{B,k+1}(h*W_k)(t), \end{cases}$$

where $E_k$ is the cumulative sum to generation $k$, and $W_k$ holds the current generation's realization (Holschneider, 2 Jan 2025).

Convergence of this infinite sum is guaranteed under the 'sub-criticality' condition $D = \mathbb E[Y] < 1$, relaxing the conventional requirement for a strictly positive lower cutoff $m_0$; micro-seismic events are fully incorporated in a mathematically well-defined manner.

4. Time-Scaling, Clock-Warping, and Practical ETAS Extensions

The clock-advance paradigm motivates direct generalizations to time-scaled (or "warped-time") ETAS models for spatiotemporal forecasting. A monotone increasing ‘warping’ $\tau(t) = \phi(t; \text{params})$ replaces actual time intervals in the triggering kernel of the conditional intensity: $$\lambda^Z(t,x,y \mid \mathcal H_t) = \tilde u(x,y) + \sum_{i: t_i < t} \kappa_{A,\alpha}(m_i) g_{c,p}\big(\tau(t)-\tau(t_i)\big) f_{D,\gamma,q}(x-x_i, y-y_i; m_i),$$ allowing aftershock clustering, background rate, and decay to be modeled under arbitrary nonlinear time scales (Das et al., 30 May 2025).

Empirically validated clock-advance (time-scale) functions include

• Calibration: $\phi_2(t; \omega) = t / \omega$
• Proportional hazards: $\phi_3(t; Z(t)) = t / Z(t)$ for covariate $Z$
• Log-linear: $\phi_4(t) = \ln t$
• Power: $\phi_5(t; \omega) = t^\omega$

These enable interpretability (e.g., calibration scale $\omega$ as a global speed-up factor or proportional hazards scale to link temporal dynamics to environmental covariates), regularize inference, and capture heterogeneous aftershock decay modes in seismic regions such as the Himalaya (Das et al., 30 May 2025).

5. Implications for Magnitude Cutoffs and Micro-Seismicity

A principal implication of the clock-advance construction is the absence of the necessity for an artificial small-magnitude threshold $m_0$. Classical ETAS requires $m\geq m_0>-\infty$ so that each earthquake has finite mean aftershocks under the branching construction. In contrast, when the background productivity-moment density satisfies $\mathbb E(Y) = \int_0^{Y_{max}} Y \psi(Y) dY < 1$ (i.e., $\alpha>b$), the cascade naturally absorbs micro-quakes as part of the background Lévy process. This ensures a mathematically well-posed model as $m_0\rightarrow -\infty$, with micro-seismicity having only a finite net clock-advance effect (Holschneider, 2 Jan 2025).

This property is particularly advantageous for modern high-resolution catalogs that register large numbers of small-magnitude events, as all data can be incorporated directly into ETAS model fitting without arbitrary preprocessing or discarding of data.

6. Simulation, Likelihood-Based Estimation, and Optimization

The clock-advance formalism introduces a new simulation and inference paradigm: one generates an initial background Lévy process, computes its local clock increments (for aftershock triggering), then iteratively samples next-generation backgrounds on these advanced clocks. This construction eliminates the need for thinning infinitesimal events.

Parameter estimation is simplified, as there is no dependence on an unobservable $m_0$, and stability reduces to ensuring $D<1$. In likelihood-based inference for the time-scaled ETAS, the standard form is: $$\ell(\beta, \theta) = \sum_{i=1}^N \ln[v(m_i)\lambda_\theta(t_i,x_i,y_i)] - \int_{t_{min}}^{t_{max}} \int_S \int_{m_0}^{\infty} v(m)\lambda_\theta(t,x,y)\,dm\,dx\,dy\,dt,$$ with the $dt$ measure unchanged by the time change but warping only affecting the triggering kernel.

Optimization techniques applied include:

• Davidon–Fletcher–Powell (DFP) quasi-Newton methods to minimize the negative log-likelihood component in the likelihood surface
• Iterative Stochastic De-clustering Method (ISDM) for background vs. triggered assignment, iterating E-steps (computing triggering probabilities) and M-steps (maximizing complete-data log-likelihood)

Both DFP and ISDM approaches benefit from clock-advance scaling, which regularizes the optimization landscape and dampens long-lag triggering, thereby improving computational efficiency (Das et al., 30 May 2025).

7. Empirical Performance and Model Comparison

Empirical studies of earthquake catalogs (e.g., Nepal, 2000–2020) using clock-advance ETAS demonstrate significant improvements in fit and forecasting capability. Key findings include:

• All time-scaled (clock-advanced) ETAS variants achieve higher log-likelihoods than unscaled models (e.g., an improvement of ≈500–700 log-units for calibration scaling $\omega=1000$)
• The ISDM approach with radially-symmetric spatial kernel and DFP optimization yields the best fit (log-likelihood $\ell=-497.87$, AIC=1011.75)
• Warped-time simplifies stochastic declustering and enhances event reclassification accuracy between background and triggered events without relying on arbitrary thresholds
• Time scaling flexibly accommodates heterogeneous aftershock decay, environmental covariates, and high-frequency micro-seismicity, supporting more precise model calibration and predictive skill (Das et al., 30 May 2025)

Model Variant	Log-likelihood	AIC
Unscaled ETAS (expo)	$-7837.83$	--
Unscaled ETAS (gamma)	$-7795.65$	--
Time-scaled ETAS (calibration, $\omega=1000$, gamma)	$-583.96$	--
ISDM + DFP (radial PDF)	$-497.87$	$1011.75$

A plausible implication is that clock-warping in ETAS models provides a natural avenue for extending earthquake forecasting methodologies to settings with temporally non-homogeneous decay, micro-seismicity, or spatially varying background rates, mitigating issues of model misspecification common under classical parametric time decay.

8. Summary and Outlook

Clock-advance ETAS offers a rigorous, multi-scale generalization of classical ETAS, synthesizing point process theory, Lévy jump processes, and flexible time-rescaling. It enables a mathematically complete incorporation of micro-seismicity, obviates the small-magnitude cutoff, and introduces efficient simulation and estimation frameworks based on infinite Markov chains and clock-warped backgrounds. Empirical results indicate substantial gains in fit and interpretability, particularly when paired with ISDM optimization for joint background/triggered decomposition (Holschneider, 2 Jan 2025, Das et al., 30 May 2025). Continued research may focus on integrating further covariate-driven time-warps, higher-dimensional Lévy backgrounds, and region-specific optimization strategies for next-generation earthquake forecasting.