Preconference Day: Summits & WorkshopsMonday, 16 November 2026 - GMT (Greenwich Mean Time, GMTZ)

A big picture look at the technology stack powering modern quant finance. How are firms leveraging machine learning, quantum computing, GPU acceleration, and cloud infrastructure to navigate complexity, scale operations, and stay ahead in an increasingly data-driven landscape?

Large Language Models mathematically manipulate language elements. But what has the philosophy of language got to say on AI and LLM?

The mathematisation of language was conceptually attempted by Gottfried Wilhelm Leibniz in the 17th century. In the 20th century, Ludwig Wittgenstein constructed alternative models of precise language and conversational language in two perplexingly different, even contradictory works and Kurt Gödel studied the features of axiomatic systems formalising a domain (integer arithmetic) paired with the broadest set of consequences possibly derived from the foundational axioms through formal logic.

Turning back to Large Language Models, what are the implications on AI and LLM? Can AI and LLM actually succeed in credibly solving problems? Is AI basically very powerful automation or does it achieve intellectual thought? And can we improve our daily use of LLM by spotting the difference?

Once code is placed inside the agent loop, the harness can do more than wrap a model. It determines what happens next, preserves relevant state, controls access to data and tools, validates intermediate results, and turns failures into corrective actions. Treating this control layer as code makes agent behavior more observable, reproducible, auditable, and secure. In finance, this creates a structured boundary around model uncertainty: probabilistic reasoning can handle ambiguity and interpretation, while execution, calculations, access controls, constraints, and validation remain explicit and verifiable.

This session connects the agent harness as an infrastructure layer with adaptive systems whose model, skill, role, and coordination choices can improve over time. The result is not a fully deterministic agent, but a system in which non-deterministic behavior is increasingly controlled through inspectable, measurable, and fully auditable execution.

We offer a new algorithm for the simulation of LSVM models, based on finance-informed learning of the leverage function. Joint work with M. Lauriere (NYU Shanghai) and T. Wagenhofer (TU Berlin & Vienna).

Traditional fixed-income risk models often rely on historical scenarios, linear factor assumptions, or manually specified shocks. These approaches can struggle when markets enter regimes that are poorly represented in the historical record, particularly when correlations shift, curves twist nonlinearly, or tail events propagate across markets in unexpected ways.

This session introduces YieldGAN, a generative modelling framework for simulating realistic forward-looking yield curve scenarios. The model learns the joint dynamics of sovereign yield curves and generates full distributions of plausible future paths, capturing nonlinear dependencies, curve-shape changes, volatility clustering, and regime-sensitive correlation structures.

We will discuss how this approach can support a range of quant workflows, including regime-break analysis, PCA/factor-structure forecasting, covariance and tail-risk estimation, stress testing, scenario generation and hedging. The session will also cover empirical validation of generated curve behaviour, including out-of-sample tests of distributional properties and correlation dynamics.

The goal is to show how synthetic market data can become a practical tool for fixed-income quants: not as a replacement for traditional models, but as a new layer for testing portfolios against plausible market environments that have not yet occurred.

• Brief introduction to RAG systems
• Metrics and automated evaluation frameworks
• Application to a use case
• Threshold calibration and monitoring
• Key conclusions