Advances in Artificial Intelligence: Implications for Capital Markets Activities: formula sheet

Key formulas

Supervised learning

Inputs (features) + known labels → predict label for new data

Use for prediction tasks: default, price direction, credit score. Needs labelled history.

Unsupervised learning

Inputs only, no labels → clusters, patterns, anomalies

Use for segmentation, anomaly detection, dimension reduction. No single 'right answer' to check against.

Reinforcement learning

Agent: state → action → reward → updated policy

Goal is to maximise cumulative reward over time. Suits sequential decisions such as execution.

Overfitting test

Training error low, out-of-sample error high → overfit

Check performance on data the model has not seen. Complex models overfit more easily.

Deep learning

Neural network with many hidden layers

A model class, not a learning type. Can be used within supervised, unsupervised or RL setups.

Implementation shortfall

Implementation shortfall = (Execution price − Decision price) × Shares for a buy order

For a sell order reverse the sign. Includes impact, spread and delay costs. Opportunity cost of unfilled shares is extra.

Quoted bid-ask spread

Spread = Ask − Bid; Half-spread = Spread ÷ 2

A market maker earns about the half-spread on each side if trades are uninformed.

Mid-price

Mid = (Bid + Ask) ÷ 2

Used as the benchmark to measure effective spread.

Effective spread

Effective spread = 2 × |Trade price − Mid|

Measures actual cost paid, which can differ from the quoted spread.

Overfitting check

Out-of-sample performance << in-sample performance signals overfitting

A rule of thumb for interpretation, not a statistical test.

Precision

Precision = TP ÷ (TP + FP)

Share of alerts that are real cases. Low precision means many false positives and heavy review workload.

Recall (detection rate)

Recall = TP ÷ (TP + FN)

Share of real cases caught. Low recall means missed fraud or abuse.

False positive rate

FPR = FP ÷ (FP + TN)

Share of legitimate cases wrongly flagged.

Accuracy

Accuracy = (TP + TN) ÷ (TP + TN + FP + FN)

Misleading when classes are imbalanced, as in fraud.

Overfitting rule

In-sample performance ≫ out-of-sample performance

A large gap signals overfitting. Test on held-out data.

SR 11-7 model risk sources

Model risk = fundamental errors in the model + incorrect or inappropriate use

Both count. A sound model used outside its intended purpose is still model risk.

Overfitting signature

In-sample error much lower than out-of-sample error

A large gap suggests the model memorised noise. Fix with out-of-sample testing, regularisation, simpler models.

Three pillars of model governance

Development and use → Validation with effective challenge → Governance, policies and controls

Validation must be independent of those who build the model.

Data quality dimensions

Accuracy, completeness, timeliness, consistency, representativeness

Use as a checklist when a question describes a data problem.

Drift types

Data drift = input distribution changes; Concept drift = input-to-output relationship changes

Both lead to degraded live performance and call for monitoring and retraining.

Technology-neutral principle

Same risk → same supervisory expectation, regardless of whether AI is used

Existing model risk, conduct and resilience rules already cover AI. Do not assume a new law is always needed.

EU AI Act risk tiers

Unacceptable (prohibited) > High-risk (strict duties) > Limited (transparency) > Minimal

Credit scoring of individuals is a typical high-risk case. Duties scale with risk.

Core governance controls

Inventory + independent validation + effective challenge + monitoring + human oversight + accountability

Use this list to build answers on how a firm should govern AI models.

Accountability rule

Accountability stays with the firm and named senior managers, never with the model or the vendor

Applies to third-party and outsourced AI as well.

Quick revision

• Machine learning learns patterns from data; deep learning uses multi-layer neural networks; generative AI creates new content.
• AI in trading supports signal generation, execution and market making, and can react at very high speed.
• In risk and compliance, AI helps with early warning, anomaly detection and surveillance of conduct and manipulation.
• Explainability is the ability to understand why a model produced an output; complex models are often less explainable.
• Poor, biased or unrepresentative training data leads to poor and biased outputs.
• Models can fail when conditions differ from the training data, as in stress periods.
• Herding risk: many firms using similar models or data may take similar positions at the same time.
• Concentration risk: reliance on a few AI or cloud providers creates third-party and operational dependence.
• Faster, correlated trading can amplify volatility and reduce liquidity under stress.
• Governance controls include independent validation, human oversight, monitoring and clear accountability.
• Existing model risk and operational resilience frameworks still apply to AI.
• Choose the answer that matches the specific risk in the scenario, not the most general one.

Common mistakes

• Treating deep learning as a fourth learning type alongside supervised, unsupervised and reinforcement. Fix: Remember that learning type describes how the model is trained; deep learning describes the model structure. A deep network can be used in any of the three.
• Calling anomaly detection or clustering supervised learning. Fix: Check whether labelled fraud cases are used. If the model only looks for unusual patterns without labels, it is unsupervised.
• Assuming AI always improves market liquidity. Fix: Remember liquidity can be fragile. Automated market makers may widen quotes or withdraw in stress.
• Confusing inventory risk with adverse selection. Fix: Inventory risk is price moves on positions held. Adverse selection is trading against better informed counterparties.
• Using accuracy to judge a fraud model. Fix: Fraud is rare, so a model flagging nothing can look highly accurate. Use precision, recall and false positive rate.
• Treating better predictive power as sufficient for approval. Fix: Supervisors also require explainability, stability, fairness and documentation. Strong accuracy alone does not meet governance needs.
• Treating model risk and AI bias as the same thing Fix: Bias is one source of model risk. Model risk also covers errors, instability, drift and misuse.
• Believing explainability tools make a black box transparent Fix: They give approximate post-hoc explanations. Validation still needs challenge, testing and sometimes a simpler model.
• Treating herding and third-party concentration as the same risk. Fix: Herding comes from similar models or data driving similar trades. Concentration comes from dependence on the same provider for a service.
• Saying AI always increases volatility. Fix: AI can improve efficiency and liquidity in normal times. The concern is that uniform behaviour can amplify stress. Use words like can and may.

Exam tips

• Questions are usually scenario-based. Identify the data and goal first, then name the method.
• Expect distractors that swap unsupervised and supervised, or treat deep learning as a separate learning type.
• When risk is asked, the usual best answers are overfitting, explainability, data quality, bias and model drift.
• For generative AI, link use cases such as summarising and coding to the hallucination and data confidentiality risks.
• Compare with traditional models in terms of flexibility versus transparency, not 'better versus worse'.
• Expect scenario MCQs where you must pair a use case with its main risk. Learn the pairs.
• Watch for stress scenarios: the correct option usually mentions liquidity withdrawal or amplification.
• Be ready for IOSCO-style governance answers: testing, monitoring, oversight and accountability.