Abstract:We propose a novel approach to marked Hawkes kernel inference which we name the moment-based neural Hawkes estimation method. Hawkes processes are fully characterized by their first and second order statistics through a Fredholm integral equation of the second kind. Using recent advances in solving partial differential equations with physics-informed neural networks, we provide a numerical procedure to solve this integral equation in high dimension. Together with an adapted training pipeline, we give a generic set of hyperparameters that produces robust results across a wide range of kernel shapes. We conduct an extensive numerical validation on simulated data. We finally propose two applications of the method to the analysis of the microstructure of cryptocurrency markets. In a first application we extract the influence of volume on the arrival rate of BTC-USD trades and in a second application we analyze the causality relationships and their directions amongst a universe of 15 cryptocurrency pairs in a centralized exchange.
What problem does this paper attempt to address?
### Problems the paper attempts to solve
The paper aims to solve the non - parametric estimation problem of high - dimensional marked Hawkes processes, especially its application in the microstructure analysis of the cryptocurrency market. Specifically, the paper proposes a moment - based neural Hawkes estimation method, which utilizes Physics - Informed Neural Networks (PINNs) to solve the Fredholm integral equation of the second kind, thereby achieving non - parametric estimation of the Hawkes process kernel function.
### Background and motivation
The Hawkes process is a point process used to model time - series of event arrivals, and these events can cluster over time. In the univariate case, the Hawkes process can exhibit self - exciting or self - inhibiting behaviors; in the multivariate case, it can exhibit mutual - exciting or mutual - inhibiting behaviors. The Hawkes process has wide applications in seismology, neuroscience, criminology, and finance, etc.
Traditional Hawkes process estimation methods usually need to specify the parametric form of the kernel function and calibrate the parameters by maximizing the likelihood function. However, this method may encounter problems such as high computational complexity and many local minima in high - dimensional cases. In addition, there may be complex dynamic behaviors in real - data, such as delay effects, which make the form of the kernel function more complex and difficult to accurately estimate by parametric methods.
### Main contributions of the paper
1. **Proposing a new non - parametric estimation method**: This method uses the latest PINNs technology to train a robust kernel function learner to solve the characteristic equation proposed by Bacry and Muzy (2016).
2. **Providing a set of general hyper - parameter and network architecture configurations**: These configurations can produce high - precision estimation results under a wide range of kernel function shapes.
3. **Application to high - frequency cryptocurrency data**: Using this estimation method, the influence of trading volume on the BTC - USD trading arrival rate is extracted, and the causal relationships and their directions among 15 cryptocurrency pairs are analyzed.
### Method overview
1. **Framework and symbol definition**: Define the intensity process and the marked process of the D - dimensional linear marked Hawkes process, and introduce the concepts of the aggregated time kernel and the aggregated marked kernel.
2. **Characteristic equation**: Based on the results of Bacry and Muzy (2016), derive the second - order statistical characteristic equation of the Hawkes process, which is a Fredholm integral equation of the second kind.
3. **Physics - Informed Neural Networks**: Utilize PINNs technology to solve the characteristic equation by training a neural network. Design a loss function combined with a weighted method of time causality to improve the training effect of the model.
4. **Sampling and training process**: Describe in detail the sampling method of data points and the training process, including steps such as random sampling, time - weight adjustment, and standardization processing.
### Application examples
1. **Influence of trading volume on trading arrival rate**: Through the analysis of high - frequency cryptocurrency data, the influence of trading volume on the BTC - USD trading arrival rate is extracted.
2. **Causal relationship analysis among cryptocurrency pairs**: Analyze the causal relationships and their directions among 15 cryptocurrency pairs, and define two ratios to quantify the directions of the causal relationships.
### Conclusion
The method proposed in the paper shows excellent performance in the non - parametric estimation of high - dimensional marked Hawkes processes, and is especially suitable for the analysis of high - frequency financial data. By applying this method, researchers can gain a deeper understanding of the microstructure and dynamic behaviors of the cryptocurrency market.