Abstract
Hawkes processes are a self-exciting stochastic process used to describe phenomena whereby past events increase the probability of the occurrence of future events. This work presents a flexible approach for modelling a variant of these, namely discrete-time Hawkes processes. Most standard models of Hawkes processes rely on a parametric form for the function describing the influence of past events, referred to as the triggering kernel. This is likely to be insufficient to capture the true excitation pattern, particularly for complex data. By utilising trans-dimensional Markov chain Monte Carlo inference techniques, our proposed model for the triggering kernel can take the form of any step function, affording significantly more flexibility than a parametric form. We first demonstrate the utility of the proposed model through a comprehensive simulation study. This includes univariate scenarios, and multivariate scenarios whereby there are multiple interacting Hawkes processes. We then apply the proposed model to several case studies: the interaction between two countries during the early to middle stages of the COVID-19 pandemic, taking Italy and France as an example, and the interaction of terrorist activity between two countries in close spatial proximity, Indonesia and the Philippines, and then within three regions of the Philippines.
Introduction
Hawkes processes serve as a critical mathematical framework for modeling "self-exciting" phenomena, where the occurrence of an event increases the probability of subsequent events. Common applications include seismic aftershocks, financial market fluctuations, and social media virality. Traditionally, these models rely on a "triggering kernel"—a function that dictates how the influence of a past event decays over time—constrained to a fixed parametric shape, such as exponential or power-law decay.
However, real-world dynamics often defy these rigid assumptions. In epidemiological contexts, for instance, the peak of excitation may be delayed by several days due to incubation periods. Furthermore, much of the existing Hawkes process literature focuses on continuous time, whereas many practical datasets (e.g., daily mortality counts) are inherently discrete. Forcing discrete data into continuous models or using overly simple kernels can lead to artificial noise and poor predictions.
Methodology: The Random Histogram Kernel
To address these limitations, we propose a non-parametric random histogram kernel for Discrete-Time Hawkes Processes (DTHPs). Rather than assuming a specific decay curve, the triggering kernel is represented as a flexible step function (a histogram).
Trans-dimensional Inference
A key innovation of this approach is the use of Reversible-Jump Markov chain Monte Carlo (RJMCMC). This trans-dimensional inference technique allows the model to treat both the number of bins in the histogram and the locations of the "change points" (the edges of the steps) as variables. By "jumping" between parameter spaces of different dimensions, the model effectively lets the data decide how complex the histogram needs to be.
Discrete-Time and Multivariate Formulation
The model is specifically formulated for discrete intervals using a Poisson likelihood. The mean of this distribution is determined by a baseline rate plus the weighted influence of past events. This framework naturally extends to multivariate scenarios, allowing for the analysis of cross-excitation between interacting processes, such as the influence of pandemic trends between neighboring countries.
Experimental Evaluation
The utility of the histogram kernel was validated through both synthetic simulations and real-world case studies:
- Simulated Data: Testing across univariate and bivariate scenarios with time-series lengths ranging from 50 to 2000 days to test the model's ability to recover known kernel shapes.
- COVID-19 Mortality: Analyzing daily death counts in France and Italy (March 2020 – May 2021) sourced from the Johns Hopkins CSSE repository.
- Terrorism Activity: Examining daily event counts in Indonesia and the Philippines (2000–2001) from the Global Terrorism Database (GTD), including a regional breakdown of the Philippines.
Key Insights and Results
Flexibility and Accuracy
Simulation results demonstrated that the histogram model achieves significantly lower Root Mean Squared Error (RMSE) than standard geometric models, particularly when the true excitation pattern is non-monotonic (e.g., a centered peak). In the COVID-19 study, the model identified a ~10-day lag in cross-country excitation between Italy and France, likely reflecting travel and incubation periods that a simple exponential decay would miss.
Sensitivity to Priors
The study also explored the impact of prior assumptions on model performance. In shorter time series (e.g., 50 days), the model exhibits sensitivity to the choice of priors. However, as data volume increases, the signal in the data dominates, leading to consistent results across different prior configurations.
Regional Dynamics
In the terrorism case study, the model revealed that Mindanao acted as a "predominant source of excitation" for other regions in the Philippines. Interestingly, cross-country excitation between Indonesia and the Philippines was found to be minimal, despite their spatial proximity, suggesting that these conflicts were largely localized during the study period.
Conclusion and Future Work
While the random histogram kernel offers unprecedented flexibility, it comes with increased computational costs. RJMCMC is resource-intensive, and scaling the model beyond three dimensions currently faces a "curse of dimensionality." Future research will focus on incorporating sparsity-inducing priors to automatically prune non-existent excitation links and treating the maximum memory length ($s_{max}$) as a parameter to be estimated rather than a fixed value.
Summary updated