---
title: Common Signals, Spurious Cross-Excitation, and Identification in Multivariate Hawkes Processes
url: https://www.ml-quant.com/papers/arxiv/2610.19428/
site: ML-Quant (https://www.ml-quant.com)
updated: 2026-10-16
license: Summaries CC BY 4.0; links go to the original sources
index: https://www.ml-quant.com/llms.txt
identifier: arXiv:2610.19428
source_url: https://arxiv.org/abs/2610.19428
featured: 2026-10-16
citations: unknown
topic: Econometrics & Forecasting
---


# Common Signals, Spurious Cross-Excitation, and Identification in Multivariate Hawkes Processes

The study characterizes how latent common signals generate spurious cross-excitation in Hawkes models and uses third-order cumulants to identify true excitation mechanisms, with applications to cryptocurrency data.

- Source: https://arxiv.org/abs/2610.19428
- Identifier: arXiv:2610.19428
- Released: 2026-10-15
- First featured: Quant Letter No. 135 (2026-10-16): https://www.ml-quant.com/issues/2026-10-16/
- Citations (Semantic Scholar): not tracked
- Published in: not yet
- Topic: Econometrics & Forecasting
- Authors: Jiwon Jung, Kiseop Lee

## Abstract (arXiv, CC0)

Multivariate Hawkes models are widely used to measure endogenous interactions in event data, but latent common signals can generate apparent excitation even without structural interaction. We characterize the resulting linear projection and construct stable, nonnegative Hawkes processes that share the mean intensity and Bartlett spectrum of common-signal models despite having different excitation mechanisms. This establishes a concrete form of second-order non-identification. We derive integrated third-order cumulants for Hawkes processes with Cox immigration and establish generic local identification of normalized structural parameters in two- and three-dimensional systems. An exact-equivalence Monte Carlo experiment demonstrates how third-order information can distinguish mechanisms with identical second-order structure, while revealing substantial finite-sample limitations under weak signals. For a restricted parametric family, we develop Whittle estimation from binned counts, establish consistency and asymptotic normality under appropriate identification conditions, and construct feasible sandwich standard errors accounting for estimated mean intensities. Simulations assess structural recovery and inference accuracy. An application to cryptocurrency transactions illustrates the sensitivity of inferred excitation to model specification and the limitations of structural interpretation based on spectral fit alone.

## Related

- [Hybrid Models for Financial Forecasting: Combining Econometric, Machine Learning, and Deep Learning Models](https://www.ml-quant.com/papers/arxiv/2505.19617/): ARIMA with SVM/LSTM: A study using econometric models, machine learning, and deep learning to predict financial trends for the S&P 500 and Bitcoin emphasizes the importance of well-constructed hybrid models for profitable trading strategies.
- [Quantile deep learning models for multi-step ahead time series prediction](https://www.ml-quant.com/papers/arxiv/2411.15674/): The article introduces a new deep learning framework for predicting multi-step time series, which improves the performance of deep learning models. It has been effectively tested on Bitcoin and Ethereum, demonstrating its ability to manage volatility and provide useful information for decision-making.
- [Enhancing causal discovery in financial networks with piecewise quantile regression](https://www.ml-quant.com/papers/arxiv/2408.12210/): The article presents a new method for building financial networks using quantile regression and a piecewise linear embedding scheme. This method uncovers intricate tail interactions in financial markets and identifies Bitcoin as the main influencer.
- [The self-exciting nature of the bid-ask spread dynamics](https://www.ml-quant.com/papers/arxiv/2303.02038/): A new State-dependent Spread Hawkes model has been proposed to forecast spread values in financial securities, incorporating the impact of the current spread state on its intensity functions.
- [Hybrid Models for Forecasting](https://www.ml-quant.com/papers/ssrn/5268691/): The research uses traditional econometric models, machine learning, and deep learning techniques to predict financial time series, using SP 500 index and Bitcoin data, and assesses the models based on forecast error metrics and trading performance indicators.
- [TKAN: Temporal Kolmogorov-Arnold Networks](https://www.ml-quant.com/papers/ssrn/4825654/): The article presents Temporal Kolomogorov-Arnold Networks (TKANs), a new neural network design that merges the benefits of Recurrent Neural Networks and Long Short-Term Memory for improved multistep time series forecasting.