Emerging Technologies: AI, Machine Learning and Automation
What this domain tests
The exam tests the difference between rules-based automation, RPA, machine learning and generative AI; the model lifecycle; the risks of bias, overfitting and unexplainability; and the EU AI Act and model risk expectations.
Core rules and concepts
- Taxonomy: deterministic rules and RPA follow explicit logic; supervised ML learns from labelled data; unsupervised ML finds structure; NLP handles text; large language models generate text and can hallucinate.
- Use cases: research summarisation, document extraction, KYC/sanctions triage, trade surveillance, reconciliation, cash forecasting, portfolio optimisation and client reporting.
- Model lifecycle: data sourcing and quality, feature design, training, independent validation, backtesting, deployment, monitoring for drift, recalibration and decommissioning.
- Model risk: overfitting, look-ahead bias, sample bias, unstable features, black-box vendor models, and unclear accountability. Explainability and human oversight are required where decisions affect clients or markets.
- EU AI Act: risk-based tiers - prohibited practices, high-risk systems, transparency obligations, general-purpose AI rules. Obligations phase in from 2025 onwards.
- Third-party models: DORA-style ICT third-party controls apply; vendor documentation and audit rights matter.
Worked example
A surveillance model reviews 10,000 alerts a month. It flags 2,000, of which 1,900 are genuine; 100 genuine cases are missed out of 2,000 true cases. Reasoning: recall = 1,900/2,000 = 95%; precision = 1,900/2,000 = 95%. Reducing the threshold would raise recall but lower precision and increase analyst workload. The exam wants the trade-off, the need to document threshold choices, and a sampling control over unflagged alerts to detect missed cases. A model with perfect backtest results but no out-of-sample validation should be treated as unproven, not as ready for production.
Traps that cost marks
Confusing RPA with AI; accepting backtest results without out-of-sample or walk-forward testing; ignoring look-ahead bias and data leakage; treating a vendor black box as validated; assuming generative AI output is factual; failing to monitor model drift after deployment.