MLARM key terms

The potential for adverse consequences from decisions based on incorrect or misused model outputs, including financial loss, reputational harm, or regulatory penalties.

A structured program of governance, controls, and validation activities that identifies, measures, mitigates, and monitors the risks arising from models across their lifecycle.

A maintained register of every model in use, recording ownership, purpose, criticality, version, dependencies, and validation status so risk can be tracked and reported.

A rating of how much a model's failure or error would affect the organization, used to set the depth of validation, monitoring, and approval required.

Independent assessment of whether a model is conceptually sound, performs as intended, and is appropriate for its use, covering design, data, testing, and limitations.

The written record of a model's purpose, methodology, data sources, assumptions, limitations, testing results, and approvals, enabling review and reproducibility by others.

A governance structure where business units own risk, an independent risk function oversees it, and internal audit provides assurance, separating model development from validation.

The end-to-end stages a model passes through, from problem definition and development to deployment, monitoring, retraining, and eventual retirement.

The controlled retirement of a model, including archiving documentation, migrating dependent processes, and confirming no downstream users rely on its outputs.

The dataset used to fit a model's parameters, whose quality, representativeness, and provenance strongly determine the model's accuracy and fairness in production.

A record of the origin, movement, and transformations of data across systems, supporting traceability, audit, error investigation, and reproducibility of model results.

Evidence about where data came from, how it was collected, and under what permissions, used to confirm lawful and appropriate use in model development.

The fitness of data for its intended use, judged by dimensions such as accuracy, completeness, consistency, timeliness, and validity.

The process of selecting, transforming, and creating input variables that make patterns easier for a model to learn, materially affecting predictive performance.

When a model learns noise and idiosyncrasies of training data so closely that it performs well in development but poorly on new, unseen data.

When a model is too simple or insufficiently trained to capture the underlying pattern, producing poor results on both training and new data.

A technique that penalizes model complexity to reduce overfitting, typically by adding a penalty term to the loss function during training.

A resampling method that splits data into multiple folds, training and testing on different subsets to estimate how well a model generalizes.

Testing a model against historical data to check whether its outputs would have been accurate, commonly used to validate risk and forecasting models.

Evaluating model or portfolio behavior under severe but plausible scenarios, such as extreme market moves or data shocks, to reveal hidden vulnerabilities.

Degradation in a model's performance over time as the relationship between inputs and outcomes, or the input distribution itself, changes.

A change in the statistical distribution of input features compared with the data used to train the model, which may undermine predictions even if relationships are stable.

A change over time in the true relationship between inputs and the target outcome, so that the same inputs should now map to different predictions.

Ongoing tracking of a deployed model's inputs, outputs, performance, and stability, with thresholds and alerts that trigger investigation or remediation.

Refreshing a model by re-fitting it on newer or expanded data to restore performance after drift or to incorporate changed business conditions.

A quantitative measure used to judge model quality, such as accuracy, precision, recall, AUC, or error rate, chosen to match the decision context.

A table of predicted versus actual classifications showing true positives, false positives, true negatives, and false negatives, the basis for many classification metrics.

Precision is the share of positive predictions that are correct; recall is the share of actual positives that are found, trading off against each other.

An incorrect positive prediction, where the model flags a case that is actually negative, often causing unnecessary cost or unfair treatment.

An incorrect negative prediction, where the model misses a case that is actually positive, often causing missed risk or harm.

The degree to which a person can understand why a model produced a given output, including the factors and logic behind the prediction.

How readily a human can grasp a model's internal mechanics; simpler models are inherently interpretable, while complex ones require post-hoc explanation techniques.

A method that assigns each feature a contribution to a prediction based on cooperative game theory, explaining individual outputs and aggregate feature importance.

A ranking of how much each input variable influences a model's predictions, useful for explanation, debugging, and reducing unnecessary inputs.

A model whose internal logic is not readily inspectable by users, such as many deep neural networks, requiring separate explanation methods to justify outputs.

Systematic error or unfair skew in model outputs or data, which can disadvantage groups and arise from unrepresentative samples, flawed labels, or proxy variables.

The principle that a model should not produce unjustified disparate outcomes across groups, operationalized through metrics such as equal opportunity or demographic parity.

A situation where a neutral-seeming practice or model produces substantially different outcomes for a protected group, a focus of fair-lending and employment scrutiny.

A characteristic such as race, sex, age, or disability that is legally shielded from discriminatory treatment and monitored in fairness testing.

Openness about how a model works, what data it uses, and its limitations, enabling stakeholders and regulators to assess and challenge its use.

A short standardized document describing a model's intended use, training data, performance across groups, and limitations, published for users and reviewers.

The policies, roles, and controls an organization uses to ensure AI systems are developed and used lawfully, ethically, and consistently with its risk appetite.

A set of principles and practices for building AI that is fair, safe, accountable, transparent, and respectful of human rights and societal values.

The study of moral questions raised by AI, including autonomy, accountability, privacy, and the just distribution of benefits and harms.

The risk that vendors, partners, or external models and data introduce failures, compliance gaps, or losses the organization must still manage.

Pre-engagement investigation of a vendor or model provider's capabilities, controls, financial stability, and compliance record before contracting or deployment.

The danger of over-reliance on a single vendor, model, or data source, so that one failure or exit disrupts many critical processes at once.

Models that produce new text, images, or code from learned patterns, raising distinct risks around hallucination, intellectual property, and data leakage.

A confident but fabricated or incorrect output from a generative model, which can mislead users if not verified against authoritative sources.

An attack where hidden or malicious instructions in input data cause a language model to ignore its intended constraints and perform unintended actions.

Deliberate manipulation of inputs to cause a model to misclassify or misbehave, used to test robustness and security of deployed systems.

A privacy principle requiring collection and use of only the personal data necessary for a stated purpose, limiting exposure if a breach occurs.

Processing personal data so that individuals can no longer be identified, reducing privacy obligations but requiring care against re-identification through combined datasets.

A lawful basis for processing personal data where the individual freely agrees, after being informed of purpose, scope, and consequences of use.

The protection of individuals' personal information through lawful collection, secure handling, and limits on use, including in model training and inference.

A published set of standards, roles, and processes for approving, documenting, validating, and monitoring models, often aligned to regulatory expectations.

Binding legal rules imposed by authorities on how AI and models must be developed, documented, and used, carrying enforcement and penalty consequences.

Classifying models or AI systems into risk levels based on impact and complexity, so that higher tiers receive stricter controls and review.

Examining how a model or business would behave under specified hypothetical conditions, used to anticipate risks that historical data does not capture.