The Price of Time Consistency Under Jump Uncertainty

When the rate or size of sudden shocks is uncertain, such as market crashes or abrupt climate transitions, a risk manager has two natural ways to hedge against not knowing the true model. The first is to fit a confidence set of jump models and take the worst one. The second is to let the worst case change over time, so that the adversary can switch to a more dangerous jump regime at any moment, depending on what has happened so far. Only the second option is "stable under pasting." Pasting stability means that splicing the first part of one plausible scenario onto the continuation of another always gives a plausible scenario. It is what guarantees time consistency: a position judged acceptable today is not judged unacceptable tomorrow without new information. The first option is simpler and less conservative, but it can contradict itself over time. This note proposes to measure that trade-off, which we call the price of time consistency, for jump-driven risk. We would build a neural solver that computes the pasting-stable worst case over jump intensity and jump-size uncertainty. We would then test on real jumpy data when the two approaches disagree, by how much, and which one gives better risk estimates and decisions.

The claim is specific and falsifiable. If the worst-case jump regime is the same at every state and time (for example, "always use the highest crash rate" for a monotone loss), the static and pasting-stable evaluations coincide, and consistency costs nothing. Divergence should appear exactly when the worst-case regime depends on the state, as with stop-losses, barriers, threshold-crossing events and optimal stopping. We expect it to be larger for jump uncertainty than for volatility uncertainty, because an adversary who controls jump timing can concentrate crashes where the payoff is most fragile. A second target is empirical evidence on an open theoretical question. A known relaxation of pasting stability for minimax stopping problems requires the densities relating one scenario to another to evolve with continuous paths. Jumps break that requirement, and whether the minimax identity then fails is unknown.

Core Claim and Positioning

The literature supplies the pieces but not this comparison. Two works establish time consistency when uncertainty is described by a family of possibly mutually singular models:

Both work only with continuous paths and uncertain volatility, not uncertain jumps. From the jump side, Dynamic Risk Measures for Processes via Backward Stochastic Differential Equations Associated with Lévy Processes builds time-consistent risk measures from backward equations driven by Lévy-process martingales, but under one fixed reference model. The dual characterizations in Coherent and convex monetary risk measures for bounded càdlàg processes say time consistency is equivalent to a concatenation property of the scenario set. None of these papers asks what imposing that property costs, or buys, when fit to data.

We deliberately do not target the most obvious move: a measure-theoretic extension of the path-space construction to jump paths (the Skorokhod space of right-continuous paths with jumps). The research file flags this as open. It is also the kind of result a pure-probability community is likely to have pursued already, and it would not tell a practitioner whether the property matters.

Our contribution is computational and empirical. It has three parts:

  • a solver for the pasting-stable jump-uncertain risk value;
  • a paired computation of the static confidence-set value on the same data;
  • decisive tests of when the two differ, and which one predicts realized losses better.

Research Program

The model class is a one- to fifty-dimensional jump-diffusion. It has drift $b$, volatility $\sigma$, jump intensity $\lambda$, and a jump-size law $F_\theta$ with parameters $\theta$. The uncertain parameters $(\lambda,\theta)$ range over a set $\Theta$ estimated from data.

We estimate $\Theta$ in two ways:

We compare two risk values for a loss functional $\xi$ of the path.

Static value. The static value is $\rho^{\mathrm{st}}_0(\xi)=\sup_{(\lambda,\theta)\in\Theta}\mathbb{E}^{\lambda,\theta}[\xi]$. Parameters are fixed over the whole horizon, so the scenario set is not pasting-stable.

Pasting-stable value. The pasting-stable value $\rho^{\mathrm{ps}}_0(\xi)=u(0,x_0)$ lets the parameters be any adapted process with values in $\Theta$. For Markovian losses, $u$ solves the robust integro-differential equation

$$\partial_t u+\sup_{(\lambda,\theta)\in\Theta}\Big\{b\cdot\nabla u+\tfrac12\mathrm{tr}(\sigma\sigma^\top\nabla^2u)+\lambda\!\int\![u(t,x+z)-u(t,x)]\,F_\theta(dz)\Big\}=0,$$

with terminal condition $u(T,\cdot)$ given by the loss. The pointwise supremum inside the equation is exactly what pasting stability buys: the adversary re-optimizes at every state.

Two gaps. We track two quantities.

  • The price of consistency is $\Delta_{\mathrm{pc}}=\rho^{\mathrm{ps}}_0-\rho^{\mathrm{st}}_0\ge 0$.
  • The inconsistency of the static rule is $\Delta_{\mathrm{tc}}=\rho^{\mathrm{st}}_0(\rho^{\mathrm{st}}_t(\xi))-\rho^{\mathrm{st}}_0(\xi)$. This is how much the static evaluator's own future re-evaluations disagree with its current one.

Solver. We use deep backward-equation training with the supremum inside the generator. The diffusion part borrows the corrected integrator from Integration Matters for Learning PDEs with Backward SDEs, which shows that naive Euler steps leave a bias that does not vanish in one-step consistency losses. The jump part is handled in one of two ways:

Whether the Euler-type bias reappears in the jump term is itself measured, not assumed.

For stopping problems we compute both orders of the minimax, $\sup_\tau\inf_Q$ and $\inf_Q\sup_\tau$, under the static set. The relaxation in Minimax theorems for American options in incomplete markets without time-consistency guarantees equality without pasting stability when the densities between scenarios have continuous paths. Under intensity uncertainty those densities jump, so a measured gap would be a numerical counterexample worth proving.

Baselines

Single fitted model (reference point). The naive reference is one fitted jump model with no uncertainty set. It is the Neural MJD fit evaluated by plain expectation and by Conditional Value-at-Risk (the average loss in the worst $\alpha$ fraction of outcomes). This is what most practitioners would ship. The strongest reported number available for the nominal model is forecasting accuracy on the S&P 500: Neural MJD reports MAE 15.4, against 34.9 for flow matching and 64.3 for a classical Merton jump diffusion. We reuse its data splits and fitting procedure. Beating it means lower tail-loss exceedance error and lower decision regret on held-out windows containing regime shifts, not better point forecasts.

Static confidence set. The strongest competitor is the static confidence-set worst case, computed by exhaustive Monte Carlo over a grid of $\Theta$. It is cheap, well understood, and uses exactly the same $\Theta$. Both methods get identical data, identical $\Theta$, and matched simulation budgets. For the solver, the budget is the same number of simulated jump paths as the Monte Carlo grid. The pasting-stable value beats it only if it improves out-of-sample coverage or stopping regret by more than the variability across bootstrap refits. If $\Delta_{\mathrm{pc}}$ is near zero, we report that consistency is free for that payoff, which is also a useful answer.

Volatility-only uncertainty. The third baseline is volatility-only uncertainty in the style of Superhedging and Dynamic Risk Measures under Volatility Uncertainty. Its volatility interval is calibrated so its one-period tail quantiles match the jump model's. This isolates whether the jump structure matters beyond fat marginals.

Discrete-time nested risk. We also run nested discrete-time risk in the style of Risk-Averse Reinforcement Learning via Dynamic Time-Consistent Risk Measures. That method mixes expectation and Conditional Value-at-Risk per step and is time-consistent by construction. It runs on the same discretized paths with its weight and tail level tuned on validation data.

Ablations. Two ablations isolate the key ingredients. The first removes the supremum, which reduces the method to a single model. The second swaps the corrected integrator for plain Euler steps, to test whether solver bias rather than the uncertainty model drives any gains.

Why not just…?

Why not just take the worst case over a static confidence set? This is the real threat. For monotone payoffs we expect it to match the pasting-stable value exactly. The proposal accepts that and uses it as a check. The open question is path-dependent losses, where the static rule can revise its own verdict with no news. Plugging $\Delta_{\mathrm{tc}}$ into the stop-loss and stopping experiments shows directly whether that inconsistency is large enough to change decisions. If it is not, the static set wins on simplicity and we will say so.

Why not just fit one good jump model and apply Conditional Value-at-Risk? Neural MJD's gains over classical jump diffusion (MAE 15.4 vs 64.3) show that the nominal model is already strong, so this could plausibly suffice in-sample. It ignores parameter uncertainty, which matters most when intensity shifts. The regime-shift backtest, where training windows exclude crash periods, is designed to settle whether the uncertainty set earns its keep.

Why not just widen the volatility band and use continuous-path theory? The Greenland analysis finds a fourth-to-second moment ratio of at least about 0.3 for temperature. That is ten times the value for dust and incompatible with any continuous diffusion, where that ratio must vanish. So a volatility band misrepresents the mechanism. Whether that misrepresentation costs anything is tested by the volatility-only baseline on gap-sensitive payoffs.

Why not just use nested discrete-time risk measures? They are time-consistent by construction, but they build ambiguity around one model rather than around jump parameters. Their answer also depends on the step size once jumps are present. We will report their sensitivity to the step size alongside ours.

Proposed Experiments

Correctness in one dimension. On synthetic Merton data with known switching intensity (the synthetic setup of Neural MJD), we solve the robust equation by finite differences to get ground truth. Solver relative $L_2$ error is compared against Euler and corrected integrators. For calibration only, the diffusion-only reference on 100-dimensional HJB is 0.049 for the corrected scheme versus 0.362 for Euler, per Integration Matters for Learning PDEs with Backward SDEs. No reported result was found on Lacuna for jump backward-equation solvers with uncertainty.

Scaling. We scale to a 10- and 50-asset basket and to the 150-dimensional jump LQR problem, where GPI-CBU reports about 9 hours on one GPU. That gives a runtime bar for adding the inner supremum.

Decisive test on S&P 500. For monotone losses, the prediction is $\Delta_{\mathrm{pc}}\approx 0$. For stop-loss and barrier losses, it is $\Delta_{\mathrm{pc}}>0$ and larger under jump uncertainty than under the matched volatility band. We report exceedance rates of realized losses and stopping regret on crash-containing held-out windows.

Decisive test on NGRIP. We ask for the risk of a warming-threshold crossing within a horizon under jump-intensity uncertainty, where the record shows shifts of up to 16°C within decades. Both the gaps and held-out coverage are reported.

Minimax check. We compute the minimax gap for American-style stopping under the static intensity set, across increasing jump intensity. A gap that grows as jumps become more frequent would be the first evidence that the continuous-density relaxation cannot extend to jumps. That result would motivate a proof.