Introduction
Hawkes Processes (HPs) are a fundamental tool in modeling "self-exciting" phenomena, where the occurrence of an event increases the likelihood of subsequent events. This behavior is prevalent in various domains, from the clustering of aftershocks in seismology to the bursty nature of high-frequency trading in finance. Traditionally, researchers have manually selected kernel shapes—such as exponential decay or power-law functions—based on domain-specific assumptions. However, this manual approach is often a bottleneck, prone to human bias, and struggles to capture complex, multi-modal behaviors.
This paper introduces a novel framework to automatically decompose these complex kernels into simpler, interpretable building blocks. By allowing for both additive and multiplicative compositions, the method not only makes these models more explainable but also significantly improves their predictive accuracy on real-world data.
Methodology: Automatic Kernel Discovery
The proposed approach follows a two-step process to find the best-fitting, most explainable kernel without human intervention:
- Discretized Kernel Estimation: The framework first estimates a numerical "blueprint" of the kernel's shape based on the empirical covariance of the event stream. This avoids the difficulties of fitting complex continuous functions directly to raw event data.
- Greedy Kernel Decomposition: The algorithm reconstructs this blueprint using a library of four base kernels: Exponential (EXP) for short-term decay, Power-Law (PWL) for long-range dependencies, Pulse (SQR) for windowed excitement, and Sinusoidal (SNS) for periodic patterns.
A key innovation is the use of both additive and multiplicative compositions. While addition captures independent triggering effects, multiplication allows the model to represent interacting behaviors, such as a periodic signal that decays over time. To ensure these composite models remain physically meaningful, the authors derived new analytical stationarity conditions to prevent the process from generating infinite event loops.
Experimental Evaluation
The framework was rigorously tested on synthetic data, financial tick data from 19 major companies, and a 50-year catalog of seismic activity.
- Financial Markets: In high-frequency trading data, composite kernels consistently outperformed single-base models. For example, in the Technology sector, the average log-likelihood improved significantly when moving from simple to composite kernels, outperforming traditional gradient-descent-based models in over 98% of tested sequences.
- Seismology: When applied to earthquake data, the algorithm automatically identified Power-Law kernels as the primary component. This finding aligns perfectly with "Omori’s Law," a cornerstone of seismology, while providing a more accurate fit than standard estimation techniques.
Discussion and Conclusion
The ability to decompose kernels into known functions provides "explainability by design." Instead of a "black-box" parameter list, researchers can interpret the model as a combination of specific physical or economic drivers. For instance, a model might reveal a strong 7-day cycle that decays slowly over months.
While a significant step forward, the authors note that deepening the decomposition can lead to overfitting. Furthermore, the current work focuses on univariate processes; extending this to multivariate Hawkes Processes—where events in one stream trigger events in another—is a necessary next step. Despite these limitations, the framework offers a robust, automated solution for kernel structure discovery in complex temporal data.
Summary updated