NuRisk: Cross-Domain Risk Inference Framework

• NuRisk is a cross-domain label for risk inference methods that incorporate latent state identification with explicit structural constraints.
• It includes techniques such as arbitrage-aware neural density extraction, innovation-tail risk estimation via nuisance autoregression, and spatio-temporal risk assessment in autonomous driving.
• Each application leverages data-driven methodologies, from generative modeling in option pricing to neural-network reinsurance control, offering practical insights into risk management.

NuRisk is a label used in recent arXiv literature for several technically distinct constructs rather than for a single unified framework. In the supplied corpus, it denotes: a risk-neutral neural generative method for extracting arbitrage-free option-implied densities; a two-step econometric procedure for estimating innovation-level tail risk under nuisance autoregression; a visual question answering benchmark for agent-level quantitative risk assessment in autonomous driving; a neural-network approach to reinsurance control under ruin-aware objectives; and a risk-neutral equivalent framework for pricing model uncertainty (Xian et al., 2024, Jurečková et al., 11 May 2026, Gao et al., 30 Sep 2025, Arandjelović et al., 2024, Wren, 19 Feb 2025). The commonality is terminological rather than formal; the underlying state variables, observables, constraints, and objective functionals differ substantially across domains.

1. Multiple meanings in current literature

Domain	NuRisk denotes	Source
Option markets	Risk-neutral generative networks for density extraction and arbitrage-aware option pricing	(Xian et al., 2024)
Time-series econometrics	Two-step nuisance-autoregression quantile procedure for innovation VaR and ES	(Jurečková et al., 11 May 2026)
Autonomous driving	VQA dataset and benchmark for agent-level spatio-temporal risk assessment	(Gao et al., 30 Sep 2025)
Insurance control	Neural-network-based reinsurance optimization under ruin-aware objectives	(Arandjelović et al., 2024)
Asset pricing under model uncertainty	Risk-neutral equivalent pricing framework with binary model risks	(Wren, 19 Feb 2025)

These usages are domain-specific. In quantitative finance, NuRisk is attached to risk-neutral valuation, model uncertainty, or actuarial control; in autonomous driving, it is the proper name of a dataset and benchmark. The supplied literature does not present a single shared formalism spanning these uses. A plausible implication is that the term has become a portable label for risk-centric methods that combine latent-state inference with explicit structural constraints.

2. Risk-neutral neural generative density extraction

In "Risk-Neutral Generative Networks" (Xian et al., 2024), NuRisk is the risk-neutral, neural generative approach embodied by RNGN. The central object is the risk-neutral density of $S_T$, recovered from European option prices through

$$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$

and, via Breeden–Litzenberger,

$$q_T(K)=e^{r(T-t)}\frac{\partial^2 C(K,T)}{\partial K^2}.$$

The construction is recast in terms of risk-neutral log-returns,

$$X_\tau=\ln\!\left(\frac{S_T}{S_t}\right), \qquad S_T=S_t e^{X_\tau},$$

with a stochastic curve on the maturity continuum driven by a single standard normal shock $Z$.

The generic generative specification is

$$X_\tau = \mu(\tau)+\sigma(\tau)\,Z\,G(Z,\tau), \qquad Z\sim\mathcal N(0,1), \quad \tau\ge 0,$$

with $X_0=0$. The paper separates the modeling of randomness from the modeling of parameter term structures. Three instantiations are emphasized. The single-maturity RN-Q model uses

$$X=\mu+\sigma Z\,G(Z), \qquad G(Z)=\big(u^{\,Z/A}+v^{-\,Z/A}+1\big),$$

with $u,v\ge 1$ and $A>0$, so that $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$0 recovers normality. The multi-maturity RN-MLP model uses

$$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$1

where $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$2, $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$3, and $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$4 are separate neural networks. The RN-DMLP model defines

$$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$5

with each $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$6 of RN-MLP form, thereby yielding a deterministic mixture of generators that can produce multi-peaked densities.

A defining feature is arbitrage-aware learning. Strike monotonicity and convexity are tied to density nonnegativity; boundary and zero-maturity conditions are enforced through $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$7; calendar monotonicity,

$$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$8

is imposed softly through penalties based on $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$9; and the martingale condition,

$$q_T(K)=e^{r(T-t)}\frac{\partial^2 C(K,T)}{\partial K^2}.$$0

is enforced through an equality penalty. Training uses mean-squared error on observed call and put prices plus arbitrage penalties,

$$q_T(K)=e^{r(T-t)}\frac{\partial^2 C(K,T)}{\partial K^2}.$$1

The architecture described as typical consists of small MLPs with two hidden layers of 32 Softplus units, a fixed learning rate such as $$q_T(K)=e^{r(T-t)}\frac{\partial^2 C(K,T)}{\partial K^2}.$$2, and large standard-normal sample sizes $$q_T(K)=e^{r(T-t)}\frac{\partial^2 C(K,T)}{\partial K^2}.$$3.

The empirical study uses S&P 500 European options from Jan 4, 1996 to Feb 28, 2023. In the single-$$q_T(K)=e^{r(T-t)}\frac{\partial^2 C(K,T)}{\partial K^2}.$$4 setting, RN-DMLP achieved testing MSE $$q_T(K)=e^{r(T-t)}\frac{\partial^2 C(K,T)}{\partial K^2}.$$5 versus DLN $$q_T(K)=e^{r(T-t)}\frac{\partial^2 C(K,T)}{\partial K^2}.$$6 and RN-MLP $$q_T(K)=e^{r(T-t)}\frac{\partial^2 C(K,T)}{\partial K^2}.$$7; on extreme moneyness, RN-DMLP MSE was $$q_T(K)=e^{r(T-t)}\frac{\partial^2 C(K,T)}{\partial K^2}.$$8 versus DLN $$q_T(K)=e^{r(T-t)}\frac{\partial^2 C(K,T)}{\partial K^2}.$$9. In the multi-$$X_\tau=\ln\!\left(\frac{S_T}{S_t}\right), \qquad S_T=S_t e^{X_\tau},$$0 setting, RN-DMLP achieved testing MSE $$X_\tau=\ln\!\left(\frac{S_T}{S_t}\right), \qquad S_T=S_t e^{X_\tau},$$1 and relative MSE $$X_\tau=\ln\!\left(\frac{S_T}{S_t}\right), \qquad S_T=S_t e^{X_\tau},$$2, versus typical baseline MSE values $$X_\tau=\ln\!\left(\frac{S_T}{S_t}\right), \qquad S_T=S_t e^{X_\tau},$$3–$$X_\tau=\ln\!\left(\frac{S_T}{S_t}\right), \qquad S_T=S_t e^{X_\tau},$$4 and relative MSE $$X_\tau=\ln\!\left(\frac{S_T}{S_t}\right), \qquad S_T=S_t e^{X_\tau},$$5–$$X_\tau=\ln\!\left(\frac{S_T}{S_t}\right), \qquad S_T=S_t e^{X_\tau},$$6. Under quote perturbations of $$X_\tau=\ln\!\left(\frac{S_T}{S_t}\right), \qquad S_T=S_t e^{X_\tau},$$7, RN-DMLP displayed the smallest deviations in density characteristics and the lowest average pricing MSE. The extracted densities showed dominant left-skew in the S&P 500 at approximately $$X_\tau=\ln\!\left(\frac{S_T}{S_t}\right), \qquad S_T=S_t e^{X_\tau},$$8 of maturities, and monthly RNM2 from RN-DMLP at $$X_\tau=\ln\!\left(\frac{S_T}{S_t}\right), \qquad S_T=S_t e^{X_\tau},$$9–35 days correlated $Z$0 with VIX. These results are attributed in the paper to flexible term structures for risk-neutral skewness and kurtosis.

3. Innovation-level tail risk under nuisance autoregression

In "Estimation of the Risk Measure under a Nuisance Autoregression" (Jurečková et al., 11 May 2026), NuRisk denotes a two-step procedure for estimating quantile-based risk measures of the unobservable innovation process in an autoregressive model with unknown nuisance parameters. The model is

$Z$1

where $Z$2 are i.i.d. with distribution $Z$3, density $Z$4, quantile function $Z$5, and assumptions including continuity and positivity of $Z$6 on its support, finite fourth moment, and a causal/stationary AR polynomial. The target risk measures are the innovation quantile

$Z$7

Value at Risk,

$Z$8

and Expected Shortfall,

$Z$9

Because $$X_\tau = \mu(\tau)+\sigma(\tau)\,Z\,G(Z,\tau), \qquad Z\sim\mathcal N(0,1), \quad \tau\ge 0,$$0, conditional risk of $$X_\tau = \mu(\tau)+\sigma(\tau)\,Z\,G(Z,\tau), \qquad Z\sim\mathcal N(0,1), \quad \tau\ge 0,$$1 is obtained by adding the linear predictor: $$X_\tau = \mu(\tau)+\sigma(\tau)\,Z\,G(Z,\tau), \qquad Z\sim\mathcal N(0,1), \quad \tau\ge 0,$$2 and similarly for ES.

The estimation problem arises because $$X_\tau = \mu(\tau)+\sigma(\tau)\,Z\,G(Z,\tau), \qquad Z\sim\mathcal N(0,1), \quad \tau\ge 0,$$3 is latent and depends on unknown $$X_\tau = \mu(\tau)+\sigma(\tau)\,Z\,G(Z,\tau), \qquad Z\sim\mathcal N(0,1), \quad \tau\ge 0,$$4. The proposed solution first estimates $$X_\tau = \mu(\tau)+\sigma(\tau)\,Z\,G(Z,\tau), \qquad Z\sim\mathcal N(0,1), \quad \tau\ge 0,$$5 by a rank-based R-estimator using the Jaeckel dispersion criterion with score

$$X_\tau = \mu(\tau)+\sigma(\tau)\,Z\,G(Z,\tau), \qquad Z\sim\mathcal N(0,1), \quad \tau\ge 0,$$6

Residuals are then formed as

$$X_\tau = \mu(\tau)+\sigma(\tau)\,Z\,G(Z,\tau), \qquad Z\sim\mathcal N(0,1), \quad \tau\ge 0,$$7

and innovation quantiles are estimated from the empirical quantiles of $$X_\tau = \mu(\tau)+\sigma(\tau)\,Z\,G(Z,\tau), \qquad Z\sim\mathcal N(0,1), \quad \tau\ge 0,$$8 or equivalently through an intercept-only autoregression quantile step using the check loss

$$X_\tau = \mu(\tau)+\sigma(\tau)\,Z\,G(Z,\tau), \qquad Z\sim\mathcal N(0,1), \quad \tau\ge 0,$$9

The paper states a reduction lemma showing that nuisance estimation does not affect the empirical indicator process at $X_0=0$0 scale, and gives the Bahadur representation

$X_0=0$1

uniformly over $X_0=0$2 in compact subsets of $X_0=0$3. Consequently,

$X_0=0$4

so the first-order variance is the same as if the innovations were observed. Feasible inference estimates $X_0=0$5 by a kernel density at the residual quantile, and confidence intervals can be built either asymptotically or by residual bootstrap. For CVaR, the estimator uses the Bassett–Koenker–Kordas check-loss characterization: $X_0=0$6

The simulation design includes AR(1) models with $X_0=0$7 and $X_0=0$8, an AR(2) model with $X_0=0$9, sample sizes $$X=\mu+\sigma Z\,G(Z), \qquad G(Z)=\big(u^{\,Z/A}+v^{-\,Z/A}+1\big),$$0, and risk levels $$X=\mu+\sigma Z\,G(Z), \qquad G(Z)=\big(u^{\,Z/A}+v^{-\,Z/A}+1\big),$$1. Innovations are Gaussian, standardized $$X=\mu+\sigma Z\,G(Z), \qquad G(Z)=\big(u^{\,Z/A}+v^{-\,Z/A}+1\big),$$2, a scale mixture $$X=\mu+\sigma Z\,G(Z), \qquad G(Z)=\big(u^{\,Z/A}+v^{-\,Z/A}+1\big),$$3, and contaminated distributions with rare large shocks. The reported finding is that at $$X=\mu+\sigma Z\,G(Z), \qquad G(Z)=\big(u^{\,Z/A}+v^{-\,Z/A}+1\big),$$4, biases are small and RMSE decreases with $$X=\mu+\sigma Z\,G(Z), \qquad G(Z)=\big(u^{\,Z/A}+v^{-\,Z/A}+1\big),$$5; at $$X=\mu+\sigma Z\,G(Z), \qquad G(Z)=\big(u^{\,Z/A}+v^{-\,Z/A}+1\big),$$6, estimation becomes harder under heavy tails and contamination, but the feasible R-based estimator remains close to the oracle that uses true $$X=\mu+\sigma Z\,G(Z), \qquad G(Z)=\big(u^{\,Z/A}+v^{-\,Z/A}+1\big),$$7. The finite-sample cost of nuisance estimation is described as negligible, and the approach is recommended when innovations may be heavy-tailed or contaminated.

4. Agent-level spatio-temporal risk assessment in autonomous driving

In "NuRisk: A Visual Question Answering Dataset for Agent-Level Risk Assessment in Autonomous Driving" (Gao et al., 30 Sep 2025), NuRisk is the proper name of a dataset and benchmark for evaluating whether multimodal LLMs can perform explicit spatio-temporal reasoning for agent-level quantitative risk assessment. The dataset comprises 2.9K scenarios and 1.1M agent-level VQA samples, assembled from 1000 Waymo scenarios, 850 nuScenes scenarios, and 1000 CommonRoad safety-critical simulations. Per-source agent-level sample counts are 617K for nuScenes, 482K for Waymo, and 64K for CommonRoad. All sources are downsampled to a uniform 2 Hz, and the visual input consists of sequential BEV images within a 30-meter radius around the ego vehicle.

The data model is specified as

$$X=\mu+\sigma Z\,G(Z), \qquad G(Z)=\big(u^{\,Z/A}+v^{-\,Z/A}+1\big),$$8

and at the agent level as

$$X=\mu+\sigma Z\,G(Z), \qquad G(Z)=\big(u^{\,Z/A}+v^{-\,Z/A}+1\big),$$9

with per-agent metrics

$u,v\ge 1$0

Risk is computed from longitudinal and lateral DTC and TTC, relative velocity, directional weighting, and thresholding, then aggregated into ordinal risk levels $u,v\ge 1$1–$u,v\ge 1$2. The paper does not provide explicit closed-form formulas for GetTTC, ComputeRisk, or CombineRisk; those components remain algorithmic abstractions.

The benchmark is structured as LLaVA-style conversation data in which the model receives sequential BEV frames and answers agent-specific questions. Three task families are emphasized: identifying the highest-risk agent at time $u,v\ge 1$3; assigning a quantitative risk level and underlying DTC/TTC values to a selected agent; and comparative temporal reasoning over the last several frames. The fine-tuning objective is a causal language modeling loss,

$u,v\ge 1$4

where $u,v\ge 1$5 is the target JSON response, $u,v\ge 1$6 is the conversation context, and $u,v\ge 1$7 are sequential BEV images.

Evaluation uses MAE, QWK, Accuracy, Precision, Recall, F1, Spatial Accuracy within 0.5 m in longitudinal and lateral directions, Temporal Accuracy within 0.5 s in longitudinal and lateral directions, and average response time. On vision-only inputs, proprietary models top out at about $u,v\ge 1$8 accuracy, while open-source models are faster but peak at about $u,v\ge 1$9 accuracy. Contextual Prompting helps, whereas CoT and ICL add little while increasing latency. When textual motion states, positions, velocities, and accelerations are added, Gemini-2.5-Flash reaches about $A>0$0 accuracy with Multi+Text and CP/CoT/ICL, and MAE about $A>0$1, although open-source models do not similarly benefit because of context-window and token-load limits. A fine-tuned NuRisk VLM agent based on Qwen2.5-VL-7B-Instruct with LoRA improves accuracy to about $A>0$2 and reduces latency by about $A>0$3 relative to the best proprietary baseline reported in the abstract; it is also reported as the only model demonstrating consistent spatio-temporal reasoning, with spatial accuracies of $A>0$4 longitudinal and $A>0$5 lateral, and temporal accuracies of $A>0$6 longitudinal and $A>0$7 lateral.

The dataset is explicitly framed as a research benchmark rather than a certified safety system. The paper notes modest absolute accuracy even after fine-tuning, abstraction in the risk aggregation functions, the effect of BEV resolution and horizon choices, and the need for rigorous validation, calibration, and fail-safes in any safety-critical deployment. It also contrasts NuRisk with prior VLM and VQA benchmarks, arguing that those benchmarks do not systematically evaluate agent-level quantitative risk with directional DTC/TTC over time in a unified BEV-sequence format.

5. Neural-network reinsurance and ruin-aware control

In the material derived from "Reinsurance with neural networks" (Arandjelović et al., 2024), NuRisk denotes a neural-network-based treatment of actuarial risk and reinsurance optimization. The setting is a finite deterministic horizon $A>0$8, discretized at times $A>0$9, with a reinsurance strategy $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$00, $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$01, $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$02. The insurer faces claims and market-dependent surplus fluctuations; in the numerical example the surplus process is a Cramér–Lundberg model perturbed by a mean-reverting Ornstein–Uhlenbeck process. Premium income follows

$$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$03

the reinsurance premium is

$$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$04

and the OU perturbation is

$$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$05

The optimization problem combines terminal wealth and finite-horizon ruin. Ruin is the discrete-time event $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$06, and the target functional is

$$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$07

with $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$08. In the numerical study the utility is exponential,

$$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$09

To obtain a differentiable training objective, the ruin indicator is replaced by the modified Gerber–Shiu surrogate

$$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$10

yielding the surrogate program

$$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$11

with $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$12. The paper states a convergence proposition: if $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$13 almost surely and $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$14 is uniformly bounded, then $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$15.

Control policies are parameterized by neural networks. The proportional algorithmic policy class is

$$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$16

where $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$17 belongs to a feedforward network class and

$$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$18

ensuring $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$19. The paper states an approximation theorem on finite $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$20: for any $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$21 and $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$22, there exists a neural-network policy $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$23 whose expected surrogate objective is within $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$24 of the supremum over admissible controls.

The numerical configuration uses $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$25, $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$26, $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$27, $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$28, $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$29, $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$30, $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$31, $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$32, $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$33, $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$34, $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$35, $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$36, and $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$37. The network has two hidden layers with tanh activations and 32 nodes per hidden layer; training uses Adam with initial learning rate $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$38, decay by $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$39 after 10 epochs without improvement, floor $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$40, early stopping after 20 epochs without improvement, 2000 batches, and batch size $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$41. On a test set of size $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$42, the paper reports that without reinsurance ($$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$43) the ruin probability is approximately $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$44, and that the surrogate loss with $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$45 matches this value closely. Numerically, time-independent feedback strategies $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$46 were sufficient. For $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$47, optimal retention is high until a surplus threshold and then drops; for $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$48, retention becomes higher once surplus exceeds approximately $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$49; for $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$50, the optimal strategy is constant. Varying $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$51 generates an approximate Pareto frontier, and all points dominate the no-reinsurance benchmark under the reported parameters.

6. Risk-neutral equivalent pricing of model uncertainty

In "The Risk-Neutral Equivalent Pricing of Model-Uncertainty" (Wren, 19 Feb 2025), NuRisk denotes a constraint-centric framework for pricing model uncertainty under a risk-neutral equivalent decomposition. The paper restricts attention to binary model risks, $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$52, and argues that viable pricing decomposes total asset-pricing into model-risk and non-model-risk components. For a bullet-payoff asset with payoff $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$53, filtration $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$54, and $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$55, the $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$56-sure conditional expectations are $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$57, and two-stage RNE aggregation takes the form

$$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$58

The total risk premium satisfies the decomposition

$$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$59

where $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$60 and

$$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$61

The model-risk-only operator is

$$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$62

and the full pricing operator is

$$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$63

Under canonical pricing and informational redundancy constraints, the model-risk premium becomes

$$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$64

The associated price-of-model-risk is

$$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$65

with

$$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$66

Positivity of $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$67 requires $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$68.

A central construct is the dynamically conserved constant

$$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$69

where $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$70. The paper interprets $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$71 as integrating ex-ante intended risk-pricing and bias. At $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$72, the price-of-model-risk is $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$73, and the gain-loss ratio equals $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$74. Bias against change is represented by $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$75, and the pair $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$76 is identified from anomaly peaks via

$$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$77

with inversion formulas

$$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$78

The framework further links model uncertainty to momentum and low-risk anomalies. Under episodic model-change risk, conditional excess-return functions are derived as

$$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$79

with peak locations and magnitudes

$$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$80

For low-risk conditioning, the peak remains at $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$81 when $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$82, but shifts to

$$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$83

when bias dominates. The paper’s stated contribution is to separate intended model-risk pricing from ex-post bias using observable anomaly patterns rather than preference recursion.

The framework is explicitly limited by assumptions: binary model uncertainty, constant economic sign of $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$84, small-increment dynamics, market completeness relative to model-risk inference, canonical pricing, episodic one-off model crises, and stationarity of $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$85 and $$C(K,T) = e^{-r(T-t)}\,\mathbb{E}_Q[(S_T-K)^+\mid\mathcal F_t],$$86 during the episode. Within those assumptions, NuRisk is presented as a practical alternative to preference-centric ambiguity models, emphasizing constraints over utility maximization and isolating model-risk pricing from within-model risk pricing.

The literature represented here therefore uses the name NuRisk for five different but structurally rigorous enterprises: arbitrage-aware generative density extraction in derivatives markets, innovation-tail estimation in autoregressive systems, spatio-temporal risk VQA in autonomous driving, neural reinsurance control in actuarial finance, and canonical risk-neutral equivalent pricing of model uncertainty. The term is thus best understood as a cross-domain label for formal risk inference under latent structure and explicit constraints, not as a single doctrine or software system.