AI and Quantitative Modelling

Definition

Quantitative modelling in risk applies probability, statistics and computation to problems where outcomes are uncertain and decisions must nonetheless be taken. Machine learning extends the toolkit but does not change the underlying obligation: a model used in a financial decision must be understood, tested and governed.

Responsible artificial intelligence in this setting means accountable ownership, documented data provenance, validated performance, meaningful human oversight and the ability to explain a decision to the person affected by it.

Why it matters

• Predictive accuracy on a historical sample is not sufficient evidence for use in pricing, credit or capital decisions.
• Complex models fail differently from simple ones: performance often degrades gradually and invisibly rather than breaking outright.
• Generative systems introduce non-determinism and prompt sensitivity into processes that previously had reproducible outputs.

Professional focus of Jonas Osman

• Choosing the simplest structure that captures the effect of interest, and justifying additional complexity on evidence.
• Bayesian methods where prior information is genuinely available and parameter uncertainty must be carried through.
• Point-process and regime-switching models for clustered, state-dependent event risk.
• Simulation design: convergence, variance reduction, dependence structure and tail resolution.

Key methodologies

• Gradient boosting and tree ensembles. Strong tabular performance paired with attribution methods and monotonicity constraints where required.
• Bayesian inference. Posterior distributions that carry parameter uncertainty into the decision rather than discarding it.
• Hawkes processes. Self-exciting intensity models for clustered event arrivals.
• Hidden Markov models. Latent regime identification with interpretable transition structure.
• Monte Carlo simulation. Distributional estimation with convergence diagnostics and explicit dependence modelling.
• Copulas and extreme-value theory. Joint tail behaviour where linear correlation is inadequate.
• Graph models. Network representations for contagion, supply-chain and counterparty structures.

Practical management applications

• Event-intensity estimation feeding pricing and accumulation monitoring.
• Scenario generation for capital, ALM and planning.
• Segmentation and early-warning analytics with human review before action.
• Model documentation and evidence packs produced as part of the build, not retrofitted.

Governance and limitations

• Explainability requirements come from the use case: a capital model, a pricing model and an internal triage tool warrant different standards.
• Bias and fairness testing is required wherever outputs affect individuals, and should be specified before deployment.
• Human oversight must be substantive — the reviewer needs the information, time and authority to disagree.
• Every AI-assisted output used in a regulated process needs a recorded owner, an audit trail and a defined failure procedure.