📊 statistics

Bayesian Non-Parametric Inference for Lévy Measures in State-Space Models

This paper introduces a novel Bayesian non-parametric framework using an Independent Gamma-scaled Dirichlet Process to infer Lévy measures for subordinators and normal variance-mean processes within linear state-space models, offering tractable inference and demonstrated effectiveness on both synthetic and high-frequency financial data.

Original authors: Bill Z. Lin, Simon Godsill

Published 2026-06-11

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Bill Z. Lin, Simon Godsill

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to understand the weather patterns of a mysterious, chaotic city. You can't see the wind directly, but you can see how trees sway, how umbrellas turn inside out, and how puddles ripple. You know the weather isn't just a smooth, predictable breeze; sometimes there are sudden, violent gusts (jumps) that don't follow the usual rules of a gentle breeze.

This paper is about building a super-smart detective kit to figure out exactly what those "gusts" look like, even when you can only see their messy effects on the ground.

Here is the breakdown of their work in everyday terms:

1. The Problem: The "Invisible Storm"

In many fields like finance (stock markets) or physics, things don't move smoothly. They jump, skip, and sometimes crash. Mathematicians call these "Lévy processes." They are great for modeling real life because they handle skewness (things leaning one way), heavy tails (rare but massive events), and discontinuities (sudden jumps).

However, there's a big catch: We can't see the storm directly.

We only see the "aftermath" (noisy data).
The storm might be happening so fast that there are infinite tiny jumps happening in a split second.
Trying to guess the exact shape of these jumps using standard math is like trying to guess the recipe of a soup by tasting a spoonful that has been mixed with mud. It's incredibly hard.

2. The Solution: A "Shape-Shifting" Detective Kit

The authors created a new Bayesian method (a way of learning from data while admitting uncertainty) to solve this. Instead of guessing a specific shape for the jumps (like "they are all bell curves"), they use a flexible, "shape-shifting" tool called the Independent Gamma-scaled Dirichlet Process (IGSDP).

The Analogy:
Imagine you have a bag of clay.

Old methods forced you to decide beforehand: "The clay must be shaped like a sphere" or "It must be a cube." If the real storm was a jagged rock, your model failed.
This new method (IGSDP) gives you a magical bag of clay that can instantly mold itself into any shape needed to fit the evidence. It doesn't force the data into a box; it lets the data tell the clay what shape to take.

3. How It Works: The "Time-Deformation" Trick

The paper focuses on a specific type of storm called a Normal Variance-Mean (NVM) process.

Think of a standard clock ticking away (time).
Now, imagine that clock is being stretched and squashed by a rubber band (the "subordinator").
When the rubber band stretches, time slows down, and the "jumps" become bigger. When it squashes, time speeds up.

The authors' method figures out two things at once:

The Rubber Band: How the time is being stretched (the subordinator).
The Storm: What the actual jumps look like once that time distortion is applied.

They use a clever mathematical trick called conjugacy. In plain English, this means their "detective kit" is designed so that every time they get a new clue, the math updates perfectly without getting messy or breaking down. It's like having a puzzle where every new piece fits perfectly into the slot without needing to force it.

4. The "Identifiability" Puzzle

One tricky part of this math is that different combinations of "rubber band strength" and "storm size" can look exactly the same. It's like saying "a small storm with a strong rubber band" looks identical to "a huge storm with a weak rubber band."

The authors solved this by fixing one part of the equation (the observation noise level). This acts like a ruler, allowing them to measure the storm accurately without the numbers getting confused.

5. Testing the Kit: From Fake Data to Real Money

They tested their detective kit in two ways:

Synthetic Data (The Simulation): They created fake data where they knew the "true" storm shape. They then tried to recover it using their method.
- Result: Even when they started with a completely wrong guess (a "bad prior"), the kit corrected itself and found the true shape. It worked even when the data was noisy (muddy) or when the storm had weird, two-peaked shapes (bimodal) that other methods missed.
Real Financial Data (The GBP/USD Forex Market): They applied this to real, high-frequency stock market data (tick-by-tick).
- Result: They used their model to predict the next move of the currency. Their "Lévy model" (which accounts for the jumps) predicted the market direction and size much better than a standard "Gaussian model" (which assumes smooth, boring weather). It was especially good at predicting sudden jumps, which standard models usually miss.

Summary

The paper presents a new, flexible mathematical framework that allows researchers to "see" the invisible, chaotic jumps in complex systems. By using a shape-shifting clay model (IGSDP) and a clever time-stretching trick, they can accurately reconstruct the hidden rules of these jumps, even when the data is messy and incomplete. They proved it works by successfully predicting real-world stock market movements better than traditional methods.

Technical Summary: Bayesian Non-Parametric Inference for Lévy Measures in State-Space Models

Problem Statement
Lévy processes are essential for modeling complex stochastic systems exhibiting skewness, heavy tails, and discontinuities, with applications ranging from finance (e.g., derivative pricing) to physics and tracking. However, inferring the underlying Lévy measures—particularly in non-parametric settings—remains a significant challenge. Existing methods largely rely on direct observations of Lévy processes or their increments. They struggle in scenarios involving discrete, noisy, and potentially partial observations, which are characteristic of linear state-space models (SSMs). Furthermore, many standard inference procedures are inapplicable to Lévy-driven stochastic differential equations (SDEs) due to intractable likelihoods and the infinite activity of many Lévy classes, where an infinite number of jumps occur within any finite time interval.

Methodology
The authors propose a novel Bayesian non-parametric framework for inferring the Lévy measures of subordinators and Normal Variance-Mean (NVM) processes within a linear state-space model. The core of the methodology involves the following components:

Independent Gamma-Scaled Dirichlet Process (IGSDP): A new flexible random measure is introduced as a prior for the subordinator Lévy measure. The IGSDP is defined as the product of a Gamma-distributed rate parameter and a Dirichlet Process (DP). This construction generalizes the well-known Gamma process (which is a special case where the shape and concentration parameters are equal) and provides tractable conditional distributions for inference.
Generative Model: The framework models the system as a linear SDE driven by an NVM process. The NVM process is constructed by time-changing a Brownian motion with a subordinator. The subordinator's Lévy measure is assigned an IGSDP prior. The model accommodates discrete, noisy observations at irregular time intervals.
Identifiability and Parameter Contour: The authors address the identifiability issues inherent in NVM processes (e.g., Generalized Hyperbolic processes) by characterizing the parameter contour. They demonstrate that while the underlying parameters ( $Q, \mu_w, \sigma_w$ ) may not be jointly identifiable, the NVM Lévy measure itself is identifiable. A practical solution is proposed: fixing the observation noise variance ( $\sigma_w^2$ ) to resolve the contour ambiguity, allowing for effective inference of the Lévy measure.
Inference Algorithm: An augmented Markov Chain Monte Carlo (MCMC) algorithm is developed to sample from the posterior distribution. Key features include:
- Collapsed Gibbs Sampling: The latent states and NVM parameters ( $\mu_w, \sigma_w^2$ ) are analytically marginalized using a Rao-Blackwellized structure, leveraging the conditional Gaussian nature of the model given the subordinator series.
- Metropolis-within-Gibbs: The subordinator Lévy measure and the jump series (sizes and times) are updated using Metropolis-Hastings steps. To improve mixing in the high-dimensional latent space, overlapping-block updates are employed.
- Hyper-parameter Conjugacy: For the Gamma process special case, the paper derives conjugate posterior updates for the concentration parameter $\alpha$ , allowing for closed-form sampling. The authors note that while the general IGSDP (with independent $\alpha$ and rate parameters) offers better flexibility, the conjugate case provides computational efficiency.

Key Contributions

Novel Prior: Introduction of the IGSDP, which extends the Gamma process and Dirichlet process, providing a flexible prior for Lévy measures with tractable conditional posteriors.
Framework for Noisy SSMs: Development of a methodology for direct inference of Lévy measures in linear state-space models under discrete, noisy, and partial observations, a setting where previous non-parametric methods fail.
Identifiability Treatment: An explicit characterization of the parameter contour for NVM processes and a practical treatment (fixing $\sigma_w^2$ ) to achieve identifiable parameterization for MCMC algorithms.
Conjugacy Results: Derivation of conjugate updates for hyper-parameters in the Gamma process special case, facilitating efficient Gibbs sampling.
Posterior Estimation: Provision of a Rao-Blackwellized posterior mean estimator for the NVM Lévy density, expressed as a mixture of Student-t kernels.

Experimental Results
The method is validated on both synthetic and real-world datasets:

Synthetic Data: The framework is tested on NVM processes driven by tempered stable (TS) subordinators and bimodal subordinators. Results show that the posterior mean estimates accurately recover the ground truth Lévy measures, even under Gaussian observation noise and with misspecified priors (e.g., using a unimodal prior to recover a bimodal measure). The method demonstrates robustness where frequentist projection and spectral methods degrade, particularly in noisy environments.
State-Space Dynamics: The approach is applied to a Langevin model (a linear SDE) driven by NVM processes. The algorithm successfully infers both the latent system states and the driving Lévy measures, even when only the position is observed with noise.
Real Financial Data: The method is applied to high-frequency, tick-level GBP/USD foreign exchange data. In a one-step and multi-step forecasting task, the Lévy-driven Langevin model outperforms a standard Gaussian Langevin model (trained via Kalman filtering) in terms of Mean Squared Error (MSE) and directional accuracy (hit rate). The Lévy model captures jumps and heavy-tailed dynamics more effectively, particularly in the immediate aftermath of market jumps.

Significance and Claims
The paper claims to significantly extend the scope of Bayesian non-parametric inference for Lévy processes by moving beyond direct observation settings to complex, noisy state-space models. By introducing the IGSDP and developing a tractable MCMC scheme, the authors provide a practical tool for modeling complex dynamics with skewness and heavy tails without relying on restrictive parametric assumptions. The work highlights that while identifying specific underlying parameters in NVM processes can be challenging, the primary goal of inferring the Lévy measure itself is achievable and yields practical benefits, as demonstrated by superior forecasting performance on financial data compared to Gaussian alternatives. The authors modestly note that future work could extend this framework to non-linear system dynamics and multivariate Lévy processes.