Risk Assessment & Management Flashcards
7 cards from real AML practice questions. Tap to flip, then mark Knew It or Still Learning — missed cards come back until you master them.
Read the first 7 Risk Assessment & Management flashcards as text
Which metric is most useful for risk assessment in a fraud detection model where false negatives are far more costly than false positives?
Answer: Recall (Sensitivity)
Recall measures the proportion of actual fraud cases correctly identified, directly capturing the cost of missing true fraudulent transactions.
A model's Expected Calibration Error (ECE) is high. What risk does this introduce in a decision-support system?
Answer: Predicted probabilities misrepresent true likelihoods, leading to misinformed decisions
High ECE means confidence scores are unreliable proxies for actual outcome probabilities, so decision thresholds calibrated on them will systematically under- or over-estimate real risk.
What is 'concept drift' and why is it a risk management concern in deployed ML systems?
Answer: A change in the statistical relationship between inputs and the target variable over time
Concept drift occurs when the underlying data-generating process changes so that previously learned input-output relationships no longer hold, degrading model reliability.
Which of the following best represents a proactive risk control in an ML model lifecycle?
Answer: Implementing automated performance monitoring with alerting thresholds before deployment
Proactive controls prevent or detect problems before they cause harm; automated monitoring with pre-set alert thresholds catches degradation before it escalates to an incident.
When performing a risk assessment for an ML model, 'inherent risk' refers to:
Answer: The raw level of risk before any mitigation controls are in place
Inherent risk is the exposure to adverse outcomes in the absence of any risk controls, providing a baseline for understanding how much mitigation is needed.
Which approach best supports 'right to explanation' requirements while managing the risk of opaque ML decisions in regulated sectors?
Answer: Applying post-hoc explainability methods such as SHAP or LIME
SHAP and LIME generate feature-level explanations for individual predictions, enabling organizations to justify automated decisions to regulators and affected individuals.
A risk review board requires a 'challenger model' alongside the production ('champion') model. What risk management purpose does this serve?
Answer: It provides a ready alternative to compare against and replace the champion if performance degrades
The champion-challenger framework maintains a validated alternative model in parallel so performance comparisons are continuous and a replacement is available without emergency development.