Many risk-management decisions are one-dimensional threshold problems. Examples are when a bank should raise capital, when an insurer should pay dividends, and when to exercise a perpetual option. The underlying quantity drifts, wiggles, and occasionally crashes downward. For such processes, called spectrally negative Lévy processes (random processes with independent increments that jump only downward), there is an unusually clean toolkit. Every quantity a decision-maker needs, such as exit probabilities, discounted occupation times, and expected payoffs at a threshold, can be written in terms of a few deterministic one-variable functions called scale functions. The optimal threshold then drops out of a smooth fit condition: the value function must join the payoff without a kink. All of this assumes the model's parameters are known. Once we admit ambiguity, say the crash intensity lies somewhere in an interval, two things are lost. One is the closed form. The other is any guarantee that today's robust assessment agrees with tomorrow's, which is the time-consistency property secured by stability under pasting: nature may switch scenarios at any random time and the result is still an admissible scenario.
The core claim is that under a pasting-stable ambiguity box over the process's drift, jump intensity and volatility, the worst case for a threshold problem is a law that switches parameters at a single state level $c$. This holds whenever a computable "generator gap" changes sign once. Because the worst-case law is a two-regime process glued at $c$, its fluctuation identities are again built from scale functions, pasted at $c$. The robust optimum is then pinned down by two first-order conditions: smooth fit at the stopping level $A$, and a second regularity condition at nature's switching level $c$. Together they give a 2×2 algebraic system in place of a nonlinear integro-differential equation. The contribution is a constructive, closed-form robust theory for one-sided jump models. It also gives a sharp, testable statement of when rectangular (pasting-stable) ambiguity gives different decisions from the naive "plug in the worst constant parameter" rule.
Where the gap is
The constructive side of this intersection is well developed, but only under a single known measure. On the Continuous and Smooth Fit Principle for Optimal Stopping Problems in Spectrally Negative Lévy Models writes threshold payoffs through the $W^{(q)}, Z^{(q)}$ scale functions. It proves that $\partial_A V_A(x)=0$ is equivalent to continuous fit for bounded-variation paths and to smooth fit for unbounded-variation paths. We reuse this "first-order condition equals regularity condition" argument for a second, adversarial player. Precautionary Measures for Credit Risk Management in Jump Models supplies our main application, a capital-raising threshold traded off against an insolvency penalty, with constant parameters. Optimality of multi-refraction control strategies in the dual model solves for two boundaries jointly through a coupled system $\Gamma(a,b)=0,\ \gamma(a,b)=0$ in refracted scale functions. It is the methodological template for our (A, c) system, and its refraction identities cover our drift-ambiguity case exactly. Periodic-decision variants, such as On the optimality of Periodic barrier strategies for a spectrally positive Lévy process, show that the required smoothness order depends on the decision structure. This warns us that the order at $c$ must be derived rather than assumed.
The ambiguity side is existential, not constructive. Constructing sublinear expectations on path space proves that the tower property of a worst-case expectation is equivalent to stability under pasting, but only on continuous paths. Superhedging and Dynamic Risk Measures under Volatility Uncertainty obtains dynamic programming via second-order backward equations, again without jumps and without free-boundary characterizations. The robust American duality in Robust Pricing and Hedging of American Options in Continuous Time is likewise restricted to continuous semimartingales.
The two Cheridito–Delbaen–Kupper papers, Coherent and convex monetary risk measures for bounded càdlàg processes and Dynamic Monetary Risk Measures for Bounded Discrete-Time Processes, supply the concatenation criterion we will verify for our scenario class. Dynamic Risk Measures for Processes via Backward Stochastic Differential Equations Associated with Lévy Processes is the only work linking Lévy dynamics to time-consistent risk measures. It works through backward-equation generators under a fixed reference measure and never produces a spatial formula or a boundary. We do not attempt the general Skorokhod-space theory that these works leave open. We work with a parametric class where pasting stability can be checked by hand, and extract formulas there.
Core Claim
Let the scenario set $\mathcal{P}$ be the laws of processes whose Lévy triplet is chosen progressively, at each instant and possibly depending on the past, from a box $\Theta$. The box is $[\mu_-,\mu_+]$ for drift, $[\lambda_-,\lambda_+]$ for the intensity of a fixed jump-size law $F$, and $[\sigma_-,\sigma_+]$ for volatility. This set is closed under pasting by construction, because concatenating two progressive parameter choices at a stopping time is again progressive. The robust value of stopping below a threshold solves an Isaacs-type equation, in which nature picks the worst generator pointwise. Write $\mathcal{L}_\theta$ for the generator under triplet $\theta$ and $V$ for the robust value. Nature's choice at state $x$ is governed by the sign of the generator gap
$$G(x) = (\mathcal{L}_{\theta_+} - \mathcal{L}_{\theta_-})V(x) = \Delta\mu\, V'(x) + \tfrac12\Delta\sigma^2\, V''(x) + \Delta\lambda \int (V(x-z)-V(x))\,F(dz),$$
where $\Delta$ denotes the width of each parameter interval. If $G$ never changes sign, the worst case is a constant parameter and robustness reduces to the static rule. Our claim concerns the single-crossing case, which arises in exactly the problems practitioners care about. Examples are two-sided exit functionals $W(x)/W(b)$, where $W$ is concave and then convex, and capital-raising problems with an insolvency penalty. In that case nature uses $\theta_-$ below a level $c$ and $\theta_+$ above it, or the reverse. The robust value is then the value under a two-regime process whose scale functions are pasted at $c$.
The two optimal levels satisfy
$$\partial_A V_{A,c}(x) = 0, \qquad \partial_c V_{A,c}(x) = 0.$$
The first condition is equivalent to classical continuous or smooth fit at $A$. We conjecture that the second is equivalent to $G(c)=0$, which corresponds to a specific regularity order at $c$. For drift ambiguity this is $V'(c)=0$. For volatility ambiguity it is $C^2$-type matching. For intensity ambiguity it is the vanishing of a nonlocal term. Pasting stability is what makes the state-feedback switch admissible. Without it, when the parameter is unknown but fixed, the answer is instead $\min_\theta V_\theta$, which is time-inconsistent.
Research Program
The first stage covers drift-only ambiguity, where the worst case is precisely a refracted Lévy process. Its exit identities and resolvents are already expressed through $W^{(q)}, Z^{(q)}$ in the multi-refraction literature. Here we prove the single-switch theorem via a verification argument. We also derive the explicit (A, c) system and show that it collapses to the Egami–Yamazaki threshold whenever $V$ is monotone.
The second stage covers intensity and volatility ambiguity. Here the two regimes differ in their jump measure or diffusion coefficient, so refraction identities no longer apply. We will derive pasted scale functions directly. The process creeps upward through $c$, which gives continuity of the upward passage. It may jump downward through $c$ with an overshoot, which yields a Volterra-type renewal equation coupling the two regimes' $W^{(q)}$ functions. We will compute the pieces in closed form by fitting $F$ with phase-type distributions, following Phase-type fitting of scale functions for spectrally negative Lévy processes and the hyperexponential construction of PHASE-TYPE APPROXIMATION OF THE GERBER-SHIU FUNCTION. With those fits the renewal equation reduces to a finite linear system.
The third stage establishes the tower property for the resulting robust value on the Skorokhod space for this parametric class. We do this by checking the concatenation criterion of the càdlàg risk-measure paper, and we state clearly that this is not a general Nutz–van Handel theorem for jumps.
Baselines
The naive reference point is the static worst-case rule. For each constant triplet in the box we compute the classical scale-function threshold, then take the parameter that minimizes value. This is what a practitioner does by stress-testing, it is time-inconsistent, and our theory says it is exact whenever $G$ keeps one sign. We run it with the same phase-type fits and report the gap in value and threshold relative to the reference solution. Beating it means showing a material threshold shift on the single-crossing problems, together with a measurable time-inconsistency of the static rule.
The strongest competitor, and our ground truth, is a direct numerical solution of the one-dimensional Isaacs equation. We use finite differences with policy iteration over nature's control and the stopping decision, with jump integrals computed by quadrature on a fine grid. No paper in the file reports results for this robust spectrally negative setting, so there is no reported result found on Lacuna for robust Lévy stopping. We run it at increasing grid resolution until it converges and give it more compute than our method. Our method beats it if it matches its converged value and its worst-case policy within phase-type tolerance, at orders of magnitude lower runtime, and stays accurate in the pure-jump $\sigma=0$ regime, where the Gerber–Shiu paper notes that discretization methods degrade near the density spike.
The third baseline is empirical dual optimisation, as in Solving optimal stopping problems under model uncertainty via empirical dual optimisation. That method reports a duality gap "frequently less than 5% of the total option value" on one-dimensional American puts. We run it on our put and credit-alarm instances with the same simulation budget. Beating it means our closed-form value lies inside its lower and upper bounds at a fraction of the cost.
The key ablation removes the switching level: we force $c$ to the boundary of the domain, so that nature plays a constant parameter. This isolates exactly the contribution of pasting-stable switching.
Why not just…?
Why not just plug in the worst constant parameter? This is the real threat, and we concede it outright for monotone or convex value functions. In those cases $G$ has one sign and our method provably returns the same answer. The proposal therefore restricts its claims to problems with an inflection in $V$. The ablation and the static baseline will show whether the resulting threshold shift is economically material or negligible. If it is negligible across the Weibull and Pareto benchmarks, that negative result is itself the finding.
Why not just solve the Isaacs equation numerically? In one dimension this works, and we use it as ground truth rather than pretend otherwise. What it does not give is an explicit system linking (A, c) to model inputs, comparative statics in the ambiguity widths, or reliable accuracy near $\sigma=0$ spikes. The runtime and pure-jump accuracy comparisons will settle whether the closed form adds anything beyond elegance.
Why not use deep optimal stopping or jump-aware neural control? Deep Optimal Stopping needs path simulation under a known measure. Deep learning for continuous-time stochastic control with jumps targets high-dimensional problems, and it reports that standard physics-informed networks struggle with jump integrals. Neither is aimed at one-dimensional problems where exact answers are available, so we treat them as out of scope rather than as competitors.
Proposed Experiments
All benchmarks are synthetic, as in the cited literature. We use the Gerber–Shiu settings: Weibull (0.6, 0.665) and Pareto (1.2, 5) claim sizes, $\sigma\in\{0,1\}$, and Monte Carlo validation with 500,000 paths. We also use the Normal, Weibull, Lognormal and Uniform jump laws from the phase-type fitting study, where approximated fluctuation identities matched simulation to the third or fourth decimal place. That accuracy sets our tolerance. The problems are the robust two-sided exit probability, the Egami–Yamazaki capital-raising threshold with ambiguous crash intensity, and a perpetual put under volatility ambiguity. Metrics are sup-norm value error and threshold error against the converged Isaacs solution, runtime, and a time-consistency score: we re-solve at simulated intermediate stopping times and count how often the decision reverses.
There are three decisive checks. First, the Isaacs solver's worst-case policy should show a single switch at the predicted $c$. If heavy-tailed Pareto jumps produce multiple switches, the single-crossing claim fails, and we extend to multi-level pasting and report the failure. Second, numerical derivatives of $V$ at $c$ should exhibit the predicted regularity order for each ambiguity type. Third, the static rule should show nonzero decision reversals, while the pasted solution shows none up to discretization error.
References
- On the Continuous and Smooth Fit Principle for Optimal Stopping Problems in Spectrally Negative Lévy Models
- Precautionary measures for credit risk management in jump models
- Optimality of multi-refraction control strategies in the dual model
- On the optimality of periodic barrier strategies for a spectrally positive Lévy process
- Constructing sublinear expectations on path space
- Superhedging and Dynamic Risk Measures under Volatility Uncertainty
- Robust Pricing and Hedging of American Options in Continuous Time
- Coherent and convex monetary risk measures for bounded càdlàg processes
- Dynamic Monetary Risk Measures for Bounded Discrete-Time Processes
- Dynamic Risk Measures for Processes via Backward Stochastic Differential Equations Associated with Lévy Processes
- Phase-type fitting of scale functions for spectrally negative Lévy processes
- PHASE-TYPE APPROXIMATION OF THE GERBER-SHIU FUNCTION
- Solving optimal stopping problems under model uncertainty via empirical dual optimisation
- Deep Optimal Stopping
- Deep learning for continuous-time stochastic control with jumps