Financial agents may behave acceptably in isolation yet create risk when deployed at scale. If many institutions use similar models, data and agent architectures, their systems may interpret events in comparable ways and take correlated actions. Feedback between agents, markets and human participants can produce outcomes that are difficult to anticipate from conventional model validation.
This course examines agentic AI as a participant in a dynamic market system. Drawing on current regulatory research, including the BIS Project Logos direction, participants will use controlled simulations to compare rules-based and LLM-based decision-makers and investigate the conditions under which collective behaviour becomes unstable.
The objective is not to teach autonomous trading. It is to equip technical and risk teams to ask better questions, design informative experiments and establish controls before agentic systems receive consequential market authority.
Learning Outcomes
Upon completion of this course, participants will be able to:
- Explain how agentic AI changes the model- and market-risk landscape
- Distinguish individual-agent performance from collective system behaviour
- Design simulations that compare heuristic and LLM-based market agents
- Identify drivers of correlated decisions and feedback loops
- Stress-test agents under changing information, constraints and market regimes
- Evaluate controls for authority, concentration and unexpected behaviour
- Interpret simulation evidence without overstating its external validity
- Define governance requirements for agentic systems in traded markets
Course Outline
From Analytical Models to Market Agents
- Predictive models, decision systems and autonomous agents
- Perception, interpretation, allocation and action loops
- Differences between advice, execution support and delegated authority
- Model risk when outputs affect the environment being modelled
- New dependencies on foundation models and shared infrastructure
- Establishing boundaries for safe experimentation
Sources of Common Behaviour
- Shared models, training data and provider infrastructure
- Similar prompts, tools, constraints and optimization objectives
- Common market narratives and information sources
- Fine-tuning, retrieval and institutional context as sources of diversity
- Model updates that alter many systems at once
- Measuring behavioural similarity rather than architectural similarity
Agent-Based Financial Market Simulation
- Purpose and limits of agent-based modelling
- Defining market environments, participants and information flows
- Heuristic, human-inspired and LLM-based portfolio agents
- Modelling constraints, transaction costs and liquidity
- Reproducibility in non-deterministic simulations
- Separating exploratory evidence from predictive claims
Designing LLM-Based Portfolio Agents
- Converting information into bounded allocation decisions
- Structured outputs and deterministic portfolio constraints
- Memory, state and sensitivity to prior interactions
- Tool access for market, risk and portfolio data
- Preventing unsupported instruments and fabricated observations
- Capturing rationales without treating them as faithful explanations
Measuring Individual and Collective Behaviour
- Allocation stability, turnover and risk-adjusted outcomes
- Dispersion, concentration and decision correlation
- Response timing and persistence after new information
- Comparing trajectories across models and random seeds
- Identifying crowded decisions and common failure modes
- Linking behavioural measures to market outcomes
Stress Scenarios and Feedback Loops
- Conflicting, ambiguous and rapidly changing information
- Volatility, liquidity deterioration and regime shifts
- Adversarial narratives and manipulated source material
- Forced constraints, margin pressure and de-risking
- Reflexive effects between agent action and market state
- Conditions that amplify or dampen correlated behaviour
Model-Risk and Systemic-Risk Controls
- Independent validation of agent objectives and tools
- Position, turnover, loss and transaction limits
- Diversity, routing and concentration controls
- Human approval and staged increases in authority
- Kill switches, circuit breakers and safe degradation
- Monitoring collective signals that are invisible at agent level
Governance and Third-Party Dependency
- Accountability for agent-led portfolio decisions
- Change control for models, prompts, tools and data
- Vendor concentration and correlated service disruption
- Evidence, audit trails and reconstruction of decisions
- Separation of research, simulation and live environments
- Regulatory and financial-stability questions for emerging systems
Practical Capstone
- Design a simulated market with rules-based and LLM-based agents
- Define behavioural and systemic-risk measures
- Run a controlled stress scenario across repeated trials
- Analyse correlated decisions and feedback effects
- Propose technical and governance controls
- Present findings, limitations and unanswered questions
Intended audience
This course is intended for senior quantitative researchers, model-risk specialists, market and operational risk professionals, quantitative developers, AI architects and technology leaders assessing agentic systems in traded markets. It is suited to participants who need to evaluate systemic behaviour rather than only individual model accuracy.
Prerequisites
Those attending this course should meet the following:
- Strong understanding of financial markets and portfolio decision-making
- Familiarity with quantitative modelling, simulation or backtesting
- Basic understanding of large language models and AI agents
- Ability to interpret statistical results; Python experience is helpful for labs