Main Conference Day TwoWednesday, 18 November 2026 - GMT (Greenwich Mean Time, GMTZ)

Discover how AI is transforming credit analysis, from unlocking new insights in ratings research to building intelligent agents and streamlining end-to-end credit workflows.

The session will cover:

• What value AI-generated features can unlock from credit and ratings research.
• How to build your own AI agent, fueled with leading data and analytics.
• An end-to-end credit workflow: from identifying risk signals to producing an auditable credit view

Gen-AI wrote the model. Who validates it? A look at where agentic techniques earn their place in risk - and where they don't.

Most institutions have run a Gen-AI experiment in risk. Very few have one in production. This session looks at what separates the two: where generative and agentic techniques genuinely improve model development, validation and hedging, and where they quietly reintroduce the risk they were meant to remove. Using Beacon Intelligence as the working example, we'll show a risk team getting from raw position to a validated, audit-ready answer without pulling the quant desk into every request, and where a human still has to sign off.

We present with DYSANOS a model framework for realistic generation of paths of daily smooth option surfaces for all strikes and expiries which are free of static arbitrage, and for which we therefore can find a change to a risk-neutral measure. DYSANOS is therefore the first production-grade ``market model`` for surfaces of European options. Our model is designed to generate entire path of daily spot and option prices for years in the future. We capture two days to expiry up to any expiry further out.

Using 1.5 billion order-level messages at microseconds resolution, we develop the first integrated framework for the identification, prediction, and trader-level attribution of endogenous systemic co-jumps. We document a synchronisation premium: simultaneous multi-asset jumps are substantially more likely than individual jumps to bear endogenous signatures. This is a classification independently corroborated by a propagation signature recovered from trade timestamps alone, under which endogenous events unfold over longer intervals, engage broader cross-sections of stocks, and arrive at smaller per-stock magnitudes than exogenous events, with mixed events consistently intermediate on every dimension. We construct Omega, a continuously computable ex-ante endogenous risk indicator integrating six microstructure features across self-excitation, phase-transition, and microstructure channels into a calibrated event-level probability whose defining property is the separation of mechanism from severity. Omega correctly distinguishes a quiet day of accumulating internal fragility from a high-stress shock-and-amplification episode of comparable realised volatility but opposite generating mechanism. A Hawkes-informed spatio-temporal graph attention network predicts endogenous co-jump incidence at the 30-minute horizon, with performance rising further during crisis windows. Ablations establish that endogenous co-jumps are fundamentally topological phenomena: removing the dual-layer trader-stock graph collapses predictive performance to near-random levels, while Omega contributes substantial incremental value. Trader attribution from two independent methodologies, pre-event behavioural entropy and learned attention weights, identifies Algo-Proprietary participants as the primary source of pre-event strategy homogenisation, with the Consumer-Banking contagion pathway recovered by both the propagation analysis and the network's learned structure.

We study whether directional macro-narrative scores extracted from news can be used to form cross-sectional signals for U.S. equity and factor portfolios.

Using 65 evergreen narratives over 2004-2025, we estimate expanding-window narrative betas and combine them with weekly changes in narrative scores to construct two related strategies: a characteristics-weighted portfolio that ranks assets by exposure to current narrative shifts, and a narrative momentum portfolio that ranks narratives by their recent score trends.

We compare a prompted large language model (LLM) sentiment score with a bag-of-words vector representation (BoW) attention measure and with a non-text benchmark based on principal components from the FRED-MD macro panel.

Across both asset universes considered, the LLM-based signals generate positive returns in both portfolio constructions, while the BoW baseline is generally weaker and the macro benchmark remains competitive.

The evidence is most naturally interpreted as showing that directional narrative measures extracted from text can complement traditional macro signals, rather than as a stand-alone replacement.

We classify return environments using turbulence measures to capture point‑in‑time relationships among CTA managers. Turbulence measures avoid rolling‑window correlations and allow decomposition into magnitude surprise and correlation surprise. Magnitude surprise reflects unusually large return movements across managers, while correlation surprise measures deviations from prevailing correlation structures. Empirically, magnitude surprise is positively related to contemporaneous and short‑term subsequent CTA returns, consistent with managed futures performing well during market stress. Magnitude surprise is also associated with elevated contemporaneous cross‑sectional dispersion, though this effect is not persistent, consistent with transitory market shocks. In contrast, correlation surprise is followed by weaker CTA performance over subsequent months, suggesting that increased disagreement across managers is detrimental to group‑level returns. Overall, the results demonstrate the value of point‑in‑time correlation measures for identifying distinct market regimes and events such as crises.

Private-firm analytics are often limited by incomplete, stale, or inconsistent financial data. Market capitalization is unavailable by definition; other financial variables such as total debt or leverage may be missing or unreliable. This paper develops a scalable framework for forecasting market capitalization and imputing missing financial quantities for private firms. The framework extends the Comparable Company Analysis by calibrating relationships on a large public-firm universe and applying them to private firms through sector, industry, region, and credit-quality information. The resulting estimates provide a internally consistent set of firm-level financial inputs that can support downstream applications such as credit-risk assessment, stress testing and private-credit portfolio analytics. The framework is designed for large-universe applications where manual valuation is infeasible, and missing values cannot simply be ignored.

We apply the framework of mixture models for Quantitative Finance tasks. In particular we base our considerations on computing conditional expectations to consider advanced and efficient regression techniques, the generation of scenarios (GenAI) including fat-tailed distributions apart from the standard Monte Carlo setting and, finally, apply adapted optimal transport techniques to the pricing and calibration of European/Bermudan/American options.

• Known modelling challenges of historic VaR and Monte Carlo VaR
• Reactivity challenge to joint projections. Persistence of joint tail behaviour in stress market regimes
• Modelling extreme loss scenarios. EVT distributions. Joint tail behaviour and choice of copula
• Can we derive a ‘universal’ copula within VaR/CVaR frameworks?
• Dissection of high dimensional empirical copula into analytical components
• Market regimes and implications for portfolio modelling