Papers

We develop a rigorous mathematical framework for the governance of systems of K self-adapting generative AI models under the principles of Model Risk Management (MRM). When multiple models share a meta-learning coupling through an interaction matrix, the per-agent Lyapunov analysis that underpins standard MRM is provably insufficient: individual agents can each satisfy their declared stability bounds while the joint system is in a regime of emergent ensemble-level drift. We formalize this gap through the Joint Lyapunov Proof (JLP)—a cryptographic and stochastic protocol that attests, without revealing proprietary weights, that the aggregate dynamics satisfy MRM Ongoing Monitoring standard at every validation epoch. Our main contributions are the following. We give a complete characterization of the infinitesimal generator of the joint quadratic Lyapunov function. We derive the exact critical coupling threshold above which the system loses mean-square stability. We prove a Noise-Floor Theorem and identify the correct target for zero-knowledge attestation. A per-epoch Succinct Non-Interactive Argument of Knowledge (SNARK) on the live weights is derived. All theoretical claims are validated against five numerical studies using a multi-agent softmax system.

Model risk management (MRM) guidance assumes a static model lifecycle, in which models are developed, independently validated, and implemented without further autonomous modification. Continually self-adapting generative AI systems — models that update their own weights during production deployment — fundamentally violate this assumption and render point-in-time validation inadequate. This paper addresses the resulting governance problem in two parts. Part I develops a rigorous telemetry architecture for such models, operating simultaneously in discrete and continuous time. We establish a Minimal Sufficient Statistic for audit purposes, construct a tamper-evident Merkle chain for discrete weight sequences, derive the appropriate continuous-time generalization via the Ito formula, and propose event-driven logging via KL divergence stopping times that is both computationally tractable and meaningful for validation. Part II asks what happens when the model provider is adversarial. A firm deploying such a model has strong incentives to conceal learning updates that would trigger mandatory validation review. We formalize this as the Model Hiding Problem and provide a systematic taxonomy of six distinct attack strategies against the Part I architecture, spanning discrete and continuous time, with a formal countermeasure for each. Together the two parts establish a dual-regime architecture in which continuous telemetry is necessary but not sufficient, narrowing—but never replacing—periodic invasive audit. The framework is model-architecture-agnostic and is designed to satisfy the three pillars of traditional MRM.

The paper places the Samuelson–Merton “util-prob” construction and preference-based state-price interpretation inside a continuous-time equivalent-martingale-measure framework. The contribution is not the existence of a marginal-utility pricing kernel, but a compensated stochastic-exponential representation of the “util-prob” density. Behavioral weighting by marginal utility becomes an equivalent martingale measure only after the drift generated by risk aversion and prudence is compensated. Under positivity, integrability, normalization, and martingale conditions, the behavioral density is generated by the Brownian kernel $-A_t\sigma_t$, where $A_t$ is Arrow–Pratt risk attitude and $\sigma_t$ is state volatility. The required compensator identifies the finite-variation term that must be removed before marginal-utility weighting defines a true martingale density. The framework preserves put–call parity under a common admissible kernel and yields a Cox–Ross risk-premium identity linking option and underlying premia through volatility exposures. Empirically, the paper maps option-implied and physical densities into a signed Arrow–Pratt decomposition, with positive curvature interpreted as risk aversion and negative curvature as risk-seeking or speculative tilt. In a panel of liquid index ETF and single-name options, arbitrage-penalized EBMM has lower pooled holdout squared-error loss than arbitrage-penalized SVI when both models receive the same total-variance positivity and butterfly-density penalties, although slice-level results remain mixed. The exercise is best read as a density-ratio admissibility test rather than a replacement for production volatility-surface estimation.

We study Lambda-quantiles, a generalisation of classical quantiles in which the constant probability level $\lambda \in [0,1]$ is replaced by a functional parameter $\Lambda \colon \mathbb{R} \to [0,1]$. We consider the general case of non-monotone $\Lambda$, which arises naturally if closure properties of the class of corresponding Lambda-quantiles with respect to inf-aggregation or with respect to mixtures are required. As preliminary results, we characterise finiteness, constancy, and what we call the attainment property known from classical quantiles. We then consider the problem of reconstructing $\Lambda$ from the values of $\Lambda$-quantiles on a suitable family of simple distributions, showing its identifiability under mild assumptions. Next, we substantially refine several results obtained in the literature on weak upper and lower semicontinuity and on the property of convexity of the level sets with respect to mixtures, obtaining in both cases almost complete characterisations without any monotonicity assumption. We then move to the case in which $\Lambda$ has bounded variation, which enables us to prove a mixture representation result: any such $\Lambda$-quantile can be rewritten as a Lambda-quantile with an increasing functional parameter, evaluated at a mixture of the original distribution with a fixed reference distribution at a fixed weight, thus reducing the complexity of the parameter from bounded variation to monotone. Finally, we introduce and study the notion of the ordinal covariance group of a risk measure, showing that in the case of a $\Lambda$-quantile it coincides with the compositional invariance group of $\Lambda$ and with a certain group of measure-preserving transformations of the signed measure associated with $\Lambda$.

The enormous growth in datasets, both in number and size, has prompted investors to adapt to new ways for assimilating information. Normatively, the approach has been to integrate such datasets into pricing formulations and assess the performance of portfolios created thereafter. However, such approaches underestimate their influence in portfolio investments by limiting their impact to pricing only. While being theoretically valid, this results in a potential sub-optimal performance in the presence of real-life decision constraints, and a blind spot for performance attribution. We start by analysing investment decisions from a knowledge perspective, which unfurls a new structure. We then propose a FinTech process termed Knowledge Optimisation that aims to integrate the influence of knowledge components that could be related to data, models, or business units that extract information. A 3-stage process, namely, decision structure, portfolio selection, and performance assessment is designed. We present an alternative to the ex-ante Sharpe Ratio, integrating a term for knowledge units. Through scenario analysis involving portfolio investment situations, we illustrate the utility. By design, the process improves the importance of knowledge in investment decisions.

Cross-correlations between financial signals are neither scale-free nor amplitude-independent: they vary with the time scale over which they are measured and with the magnitude of the fluctuations that dominate the average. We exploit this structure to construct a portfolio allocation model in which the risk functional is the signed fluctuation function of multifractal cross-correlation analysis (MFCCA), indexed by a scale $s$ and a fluctuation order $q$. Unlike MFDCCA-type criteria, which rectify local detrended covariances before aggregation, MFCCA retains their sign, so that co-moving and counter-moving components contribute to risk with opposite signs; for $q=2$ the resulting quadratic form coincides with the detrended fluctuation function of the portfolio series itself, recovering the mean–variance criterion as a scale-dependent limit. Using two-component ARFIMA and Markov-switching multifractal processes, we show that prescribed multiscale and multifractal dependence is transmitted into the optimal weights, and that sign preservation contributes more to the reduction of tail risk than aggregation over fluctuation orders. Applied to financial multi-assets, the criterion lowers drawdown, Value-at-Risk, and expected shortfall relative to the mean–variance benchmark at every required return, in and out of sample, without any loss in realized portfolio return. The construction maps signed multiscale interaction structures onto resource-allocation decisions, and applies to any complex system whose components interact across heterogeneous scales with amplitude-dependent coupling.

The impact of web datasets on market prices has suggested the development of new sources of information, such as social media and web portals, indicating the possibility of an emergent phenomenon. We propose a defining perspective, termed open information, that adds to the existing types of public and private information. We demonstrate their existence and justify material significance for pricing. In this respect, we present statistical hypotheses to test for a web dataset, GDELT, integrated for carbon pricing, that is represented by EU Allowance spot prices. Tests are designed with VAR and GARCH-X formulations, and return forecasting. The outcomes cannot rule out the material existence of open information. The result is significant for providing a conceptual basis to integrate a vast number of web datasets as alternative data in investment decisions.

Forecasting systems used in the Comprehensive Capital Analysis and Review (CCAR) and Current Expected Credit Losses (CECL) processes combine portfolio data, macroeconomic scenarios, model specifications, business assump- tions, and management adjustments. When the forecast changes from one run to the next, practitioners need an attribu- tion that reconciles to the total change without depending on an arbitrary sequence of input replacements. This paper formulates forecast-gap attribution as a cooperative game and examines several approaches: the exact Shapley value, hierarchical or nested Shapley values, Integrated Gradients, Gradient SHAP, Permutation SHAP, and Kernel SHAP. We compare their allocation rules, computational costs, implementation requirements, and limitations in production forecasting systems. The analysis provides a practical framework for choosing an attribution method according to the number and type of inputs, the feasibility of hybrid forecast runs, and the need for interpretability, reproducibility, and governance.

Risk-aware Q-learning (RaQL) provides a model-free, two-timescale estimator for dynamic risk objectives, but its finite-budget behavior remains fragile: fixed inner-loop hyperparameters can produce unstable value estimates, persistent Bellman residuals, and inefficient sample reuse. This paper proposes an adaptive training controller for Conditional Value-at-Risk (CVaR) RaQL and evaluates it on a daily Bitcoin trading task. The controller preserves the original CVaR estimator and Bellman fixed point; instead, it redesigns the training procedure through six coordinated mechanisms: per-cell inner-step sizing, outer-rate-matched decay synchronization, a short early correction for the VaR-like inner variable, a coverage-first-then-greedy sample allocation rule, progressive suffix aggregation of mature inner estimates, and data-driven calibration of key scales from online-observable quantities. Across 20 random seeds and 856,000 inner-transition samples, the controller reduces the mean empirical CVaR Bellman residual by approximately 85% relative to the fixed-parameter baseline (MeanBEQ: 1.2202 to 0.1854; MeanBEV: 1.1624 to 0.0535) and maintains stability across CVaR levels, discount factors, and training budgets. On the chronological out-of-sample test set, the learned policy attains a Sharpe ratio of 0.9281 with a maximum drawdown of 6.46% after transaction costs. Although buy-and-hold yields a higher cumulative return (35.43% vs. 23.61%), the adaptive policy achieves far lower volatility (9.57% vs. 47.93%), drawdown, and CVaR loss. These results demonstrate that adaptive finite-budget training design, applied solely to the training procedure without altering the risk objective, can materially improve the reliability and risk-adjusted performance of risk-aware Q-learning in financial applications.

We revisit the “collapse to the mean” phenomenon, which refers to mild structural conditions, such as local linearity, that force a law-invariant functional $\ph$ defined on finite-mean random variables to depend solely on the expectation of its argument $X$, and not on any other distributional feature. Starting from a concise characterisation of the convex order, our simplified approach unifies and extends existing results without assuming the functional to be convex or monotone in the almost-sure order, and clarifies the conceptual foundations of the “collapse to the mean" phenomenon. In addition, we establish a new “dual collapse” result for quasi-star-shaped functionals.

We introduce a framework for preference-robust decision making when preferences over risk are modelled through generalised distortion risk measures. Unlike distributional robustness, our approach addresses ambiguity in the risk functional itself. We construct ambiguity sets on distortion (weight) functions using the Wasserstein distance and Bregman divergences, and derive closed-form expressions for the worst- and best-case distortion risk measures. We further extend the framework to rank-dependent utility, yielding preference-robust behavioural models. In particular, rank-dependent utility appears as a robustification of the expected utility model, yielding a novel way to address the Allais paradox.

Agentic AI is gaining acceptance in asset management, but governance has not kept pace: 88% of surveyed finance professionals report no operational governance framework for agentic AI, and only 24 of 75 large U.S. money managers disclosing AI use in Form ADV filings report a formal governance policy. We argue this gap is architectural: governance built for static validation does not survive continuously retrained agentic policies. We propose a four-layer framework (Policy, Engineering, Composition, Systemic) grounded in two distinct kinds of evidence, kept explicitly separate: two calibrated synthetic illustrations (a regret-covariance drift monitor; a crowding simulation showing joint drawdown risk rising from 39.2% to 79.3%), and three real, documented cases (a deployed LLM-embedding trading strategy, a \$45 billion discretionary fund's forced-deleveraging blowup, and a tribunal ruling holding an airline liable for its chatbot). The synthetic examples demonstrate computability from observable data; the cases demonstrate that the failure modes are not hypothetical. We provide a 90-day implementation sequence spanning trading and payments/customer-facing systems.

We formulate an over-the-counter (OTC) market-making problem in which request-for-quote (RFQ) arrivals are modelled by general Hawkes kernels and fills are controlled thinnings of the exogenous request flow. The modelling choice is motivated by spot-FX RFQ data: after filtering and transforming to seasonality-adjusted RFQ activity time, two-way activity in major currency pairs exhibits large fitted branching ratios and multi-scale persistence. For general Hawkes kernels the control problem is path-dependent: the relevant state contains the order-flow history, or equivalently the forward curve of conditional future RFQ intensities. Exact Markovian lifting is available for exponential kernels, but it becomes high-dimensional for mixtures of exponentials and impractical for long-memory kernels. We therefore develop a hierarchy of Volterra-Riccati approximations. The first level replaces random future request flow by its conditional Volterra forecast; the second adds a covariance correction for intensity uncertainty; the third updates the quote rule with the realized Hawkes memory, or equivalently with the post-request conditional forecast curve. The approximation hierarchy is validated in an exponential Hawkes benchmark, where the exact lifted HJB can be solved numerically. The state-feedback Volterra-Riccati policy closely tracks the exact benchmark, especially in directional regimes, while a memory-free Poisson policy suffers substantial regret. We then apply the same state-feedback rule to a power-law-like RFQ memory model. A directional RFQ burst changes the conditional forecast of future flow and is converted by the continuation-value shadow price into a persistent quote skew. The resulting endogenous OTC quote impact inherits the long-memory decay of the RFQ forecast response and improves inventory and P&L risk control relative to a no-conditioning Poisson benchmark.