11.1_Chapter_Summaries - ravkorsurv/kor-ai-core GitHub Wiki

11.1 Chapter Summaries – Risk Assessment & Decision Analysis with Bayesian Networks

This section summarizes each chapter of Fenton & Neil’s Risk Assessment and Decision Analysis with Bayesian Networks and aligns it with Kor.ai’s design of Bayesian risk models for market abuse detection.

Chapter-by-Chapter Summary & Relevance

Chapter 1 – Introduction

  • Key Idea: Traditional risk assessment models are limited by reliance on historical frequency, lack of transparency, and poor causal reasoning.
  • Kor.ai Relevance: Validates use of probabilistic models for risk scoring. Supports abandoning deterministic alert thresholds.

Chapter 2 – Problems with Traditional Risk Assessment

  • Key Idea: Risk matrices, scorecards, and control frameworks are flawed due to assumptions of independence and lack of causality.
  • Kor.ai Relevance: Reinforces our approach of using cause-and-effect modeling (e.g., intent → action → outcome) for insider dealing or spoofing.

Chapter 3 – Probability Basics

  • Key Idea: Focus on conditional probability, belief updates, and Bayes' Theorem. Probabilities reflect belief, not just frequency.
  • Kor.ai Relevance: We model expert priors on access to information or intent where limited labelled data exists.

Chapter 4 – Common Fallacies

  • Key Idea: Base rate neglect, conjunction fallacy, and correlation ≠ causation are common in naive statistical approaches.
  • Kor.ai Relevance: Avoids false positives from coincidental trade behavior; instead models causal likelihood (e.g. negative news + access = higher risk).

Chapter 5 – Uncertainty & Expert Judgement

  • Key Idea: How to elicit probabilities from SMEs when data is unavailable. Use ranges, rankings, or visual probability tools.
  • Kor.ai Relevance: CPTs for nodes like “Intent to Manipulate” are bootstrapped from compliance expert input in MVP stage.

Chapter 6 – Anatomy of a Bayesian Network

  • Key Idea: Describes nodes, links, conditional probability tables (CPTs), and belief propagation.
  • Kor.ai Relevance: Directly maps to how we define our insider dealing and spoofing BN structures. Each alert type is a subgraph.

Chapters 7–10 – Building Bayesian Networks

  • Key Idea: Step-by-step guide on node design, causal direction, “explaining away” (abduction), and use of hidden nodes.
  • Kor.ai Relevance: Encourages use of latent factors (e.g., intent, opportunity). Spoofing model uses “order intent” → “quote behavior” → “price movement”.

Chapter 11 – Noisy OR and Ranked Nodes

  • Key Idea: Simplifies CPTs when multiple causes lead to a single effect; supports ordinal variables.
  • Kor.ai Relevance: Combines soft signals (e.g. price spike + negative news) to explain alert scoring without exponential CPT complexity.

Chapters 12–13 – Parameter and Structure Learning

  • Key Idea: Use real data to learn CPTs and network structure using Expectation Maximization (EM) or scoring methods.