Standard models of change often rely on the continuous path assumption. Whether tracking a stock price or a physical system, we typically assume that to get from point A to point B, the system must pass through every point in between. Mathematically, this is represented by Brownian motion. However, as Peng et al. and Miao et al. argue, this assumption fails to capture the "shocks" inherent in real-world systems—sudden market crashes, overnight price gaps, or instant insurance claims.
Jump-diffusion modeling addresses this by augmenting continuous diffusion with discrete, stochastic jumps. This framework allows for a more realistic representation of systems where "physics" can change instantly.
The Intuition of Discontinuity
In a pure diffusion model, the probability of a large change over a very short time interval is nearly zero. In reality, news releases or liquidity crises cause prices to "jump" across values without visiting them. Yang et al. (2023) demonstrate that market reactions to influential news are better modeled as discontinuous jumps within a stochastic differential equation (SDE) rather than smooth trend lines.
This distinction is not just theoretical; it changes how we calculate risk. Huang et al. show that in highly volatile markets like cryptocurrency, the intensity of these jumps can be an order of magnitude higher than in traditional equities. By incorporating Poisson-driven jump components, the authors can represent "fat-tailed" distributions—scenarios where extreme events are far more likely than a standard Gaussian model would predict.
Directional Shocks and Spectrally One-Sided Processes
A powerful subset of this research focuses on spectrally one-sided Lévy processes. These models are designed for systems where jumps occur primarily in one direction:
- Spectrally Negative Processes: These allow for sudden downward jumps while maintaining continuous upward paths. Yamazaki et al. (2012) use this to model "jump-to-default" risks in credit swaps, where a firm's value might erode slowly but collapse instantly. Similarly, Yamazaki (2017) applies this to insurance surplus, where premiums flow in continuously but claims arrive as sudden, discrete hits.
- Spectrally Positive Processes: These model systems with gradual decay and sudden windfalls. Avanzi et al. and Czarna et al. apply this to high-growth startups or mining operations, where continuous operational costs are punctuated by random, massive revenue jumps.
By restricting the direction of jumps, researchers can use "scale functions" to derive closed-form solutions for complex problems like optimal dividend payouts or bankruptcy thresholds, which would be intractable in more general models.
The Challenge of the Overshoot
One reason jump models are mathematically difficult is the overshoot problem. In a continuous model, a process hits a boundary (like a stop-loss trigger) exactly at the boundary value. In a jump model, the process can "jump over" the threshold.
Yamazaki et al. (2014) explain that this requires more sophisticated boundary conditions. For example, when deciding the optimal time to exercise an option, one must account for the fact that the asset price might land deep in the "exercise zone" in a single step. Pérez et al. (2017) further complicate this by exploring "periodic exercise," where opportunities to act only occur at discrete, random intervals, reflecting the reality that investors cannot monitor markets with infinite frequency.
Bridging to Machine Learning
Recent work has begun integrating these discontinuous dynamics into deep learning. Traditional Neural SDEs often assume continuous Brownian motion for ease of gradient computation. However, Kidger et al. and Wildner et al. identify the incorporation of jump diffusions into deterministic inference frameworks as a critical frontier.
Models like LEGEND (Wang et al.) and STrader (Yang et al. (2022)) represent the first steps toward this, using generative frameworks and normalizing flows to handle the non-linear, discontinuous transitions found in aggregate population data and intraday portfolio management. These approaches move beyond treating jumps as "noise" to be filtered, instead treating them as the primary signal of system state changes.
Key Foundational Work
Utility maximization in incomplete markets (2005) — contributed 2 concept elements to this direction.