What will I learn in chapter 7 of Financial AI in Practice?

The introduction frames fraud as a digital, evolving threat and AI as both a defense and a tool for fraudsters. After reading it you should be able to: Explain how financial theft has shifted from physical cash to exploiting digital systems, and why the growth of online transactions creates more opportunities for fraud. Explain why AI complicates fraud defense: institutions can use it to detect and prevent fraud, while adversaries can use it to make fraud more convincing or operate at greater scale. Explain why fraud detection matters beyond preventing direct financial losses, and use the scale of prevention investment to illustrate its urgency. Describe how digital transaction growth and increasingly sophisticated attackers make fraud harder to prevent, including the role of GenAI. Explain the tension between protecting customers from fraud and keeping legitimate transactions convenient. Distinguish transaction-based, identity-based, and application-based fraud by what is abused and when, using an example of each. Distinguish internal or employee fraud, market manipulation fraud, and social engineering-based fraud by whom or what each targets, using an example of each. Explain why fraud detection is an active arms race and how fraudsters’ adaptation can make a previously effective detector obsolete.

What does section 7.1, The criticality of fraud detection and prevention, cover?

The section explains why digital growth, evolving fraud, regulation, and customer expectations make effective fraud detection a strategic necessity for financial institutions. Explain why fraud detection matters beyond preventing direct financial losses, and use the scale of prevention investment to illustrate its urgency. Describe how digital transaction growth and increasingly sophisticated attackers make fraud harder to prevent, including the role of GenAI. Explain the tension between protecting customers from fraud and keeping legitimate transactions convenient.

What does section 7.3, The distinctive nature of financial fraud, cover?

The section explains why financial fraud is a distinctive AI challenge and why anomaly detection is useful but insufficient on its own. Explain why fraud detection is an active arms race and how fraudsters’ adaptation can make a previously effective detector obsolete. Describe how the rarity of fraudulent transactions makes them difficult for traditional ML models to identify. Distinguish the costs of missing fraud from the costs of blocking a legitimate transaction, and explain why detection involves a tradeoff.

What does section 7.4.2, Data asset layer for AI-powered fraud detection, cover?

This section explains what fraud detection needs from the data asset layer: behavioral and relational data, fraud-specific features, and pipelines fast enough to act while fraud is unfolding. Distinguish the data needed for fraud detection from the data emphasized in credit scoring, while noting that both use the same basic data asset layer functions. Describe the main kinds of data gathered for fraud detection and explain why connections across accounts and entities are particularly valuable. Describe the kinds of features engineered from fraud data, including rapid activity, deviations from a user’s habits, suspicious links, and session or device trust.

What does section 7.4.3, Modeling layer: AI techniques for fraud detection, cover?

This section explains how different fraud-detection models address known fraud, novel anomalies, connected fraud rings, and scarce labels, then positions LLMs within a layered fraud-defense system. Explain how supervised models learn from labeled fraud and legitimate cases, why they work well for known patterns, and how sparse, imbalanced labels and new tactics limit them. Describe how unsupervised methods flag departures from predominantly legitimate behavior and why their ability to surface new fraud comes with false positives and a changing baseline. Explain how graph models represent entities and relationships to expose coordinated fraud that individual-transaction models may miss, and identify their practical complexity.

What should I know before reading chapter 7?

Putting it all together: The four-layer framework in production (chapter 6): The earlier application of the four-layer framework to scoring provides a contrast with this adversarial fraud setting.; Core building blocks of financial AI applications (chapter 1): Introduces the four-layer framework for financial AI applications.; A domain-driven four-layer framework (chapter 4): Applies the framework to credit scoring.; Data asset layer (chapter 4): The data asset layer in the credit scoring framework.

Chapter 7: Understanding financial fraud and foundational detection strategies

What you will learn in chapter 7 of Financial AI in Practice

The introduction frames fraud as a digital, evolving threat and AI as both a defense and a tool for fraudsters.

After this chapter you will be able to

Shift to digital fraud. Explain how financial theft has shifted from physical cash to exploiting digital systems, and why the growth of online transactions creates more opportunities for fraud.
AI as a double-edged sword. Explain why AI complicates fraud defense: institutions can use it to detect and prevent fraud, while adversaries can use it to make fraud more convincing or operate at greater scale.
Fraud prevention as strategic imperative. Explain why fraud detection matters beyond preventing direct financial losses, and use the scale of prevention investment to illustrate its urgency.
Expanding digital fraud exposure. Describe how digital transaction growth and increasingly sophisticated attackers make fraud harder to prevent, including the role of GenAI.
Customer security friction paradox. Explain the tension between protecting customers from fraud and keeping legitimate transactions convenient.
Payment identity and application vectors. Distinguish transaction-based, identity-based, and application-based fraud by what is abused and when, using an example of each.
Insider market and human vectors. Distinguish internal or employee fraud, market manipulation fraud, and social engineering-based fraud by whom or what each targets, using an example of each.
Adversarial behavior and drift. Explain why fraud detection is an active arms race and how fraudsters’ adaptation can make a previously effective detector obsolete.
Opposing error costs. Distinguish the costs of missing fraud from the costs of blocking a legitimate transaction, and explain why detection involves a tradeoff.
Anomalies versus intentional fraud. Distinguish anomaly detection from fraud detection and explain why anomaly scores alone cannot establish fraud.
Four-layer fraud detection lens. Explain how the four-layer framework provides a lens for understanding the architecture of fraud prevention systems across financial institutions, and name its layers.
Fraud rules and rule engines. Explain how human-defined if-then conditions flag suspicious activity and how a rule engine combines and applies rules to incoming transactions.
Rule performance tradeoffs. Describe how analysts test proposed rules on historical transactions and weigh fraud caught against legitimate transactions incorrectly flagged.
Shared links in transaction graphs. Explain how representing users and merchants as nodes connected by transactions can expose a possible shared point of compromise among chargeback cases.
Foundational methods across four layers. Trace how both the rule example and the chargeback investigation move from data and simple analysis through strategic decisions to operational actions.
Fraud data versus credit data. Distinguish the data needed for fraud detection from the data emphasized in credit scoring, while noting that both use the same basic data asset layer functions.
Fraud evidence and connections. Describe the main kinds of data gathered for fraud detection and explain why connections across accounts and entities are particularly valuable.
Fraud specific feature signals. Describe the kinds of features engineered from fraud data, including rapid activity, deviations from a user’s habits, suspicious links, and session or device trust.
Real time feature pipeline. Trace how streamed fraud events become features and reach models, and explain why the pipeline must respond quickly.
Supervised fraud pattern detection. Explain how supervised models learn from labeled fraud and legitimate cases, why they work well for known patterns, and how sparse, imbalanced labels and new tactics limit them.
Unsupervised anomaly detection tradeoffs. Describe how unsupervised methods flag departures from predominantly legitimate behavior and why their ability to surface new fraud comes with false positives and a changing baseline.
Relational fraud graph analysis. Explain how graph models represent entities and relationships to expose coordinated fraud that individual-transaction models may miss, and identify their practical complexity.
LLM investigation and grounding. Describe the fraud-management tasks suited to LLMs and how a managed RAG pipeline grounds their case summaries and analysis in controlled sources.
Layered real-time fraud defense. Justify separating fast transaction blocking from asynchronous LLM analysis, and describe how cost, privacy, and audit requirements influence LLM deployment.
Risk tiers and actions. Explain how the strategy and monitoring layer converts model outputs into actions by setting score thresholds for low-, medium-, and high-risk cases, while accounting for institutional constraints.
Policy rules and lists. Describe how explicit policies and maintained watchlists supplement score-based fraud decisions.
Analyst review feedback loop. Trace how alerts reach human analysts and how investigation outcomes improve fraud defenses.
Performance monitoring and governance. Explain what the team monitors and why ongoing validation and governance are needed.
Operational role and integration. Explain how the application layer turns upstream fraud intelligence into actions involving customers, analysts, and banking systems, and why integration matters.
Real-time transaction decisions. Trace how an initiated payment can lead to an approval, decline, or authentication challenge.
Modular fraud system resilience. Explain why separating the data asset, modeling, strategy and monitoring, and application layers makes a fraud system more resilient than a monolithic design. Describe how one component can change without disrupting subsecond payment authorization.

liveBook manning.com

Read chapter 7 on liveBook Get the book at manning.com

What it builds on

Putting it all together: The four-layer framework in production (chapter 6) · The earlier application of the four-layer framework to scoring provides a contrast with this adversarial fraud setting.
Core building blocks of financial AI applications (chapter 1) · Introduces the four-layer framework for financial AI applications.
A domain-driven four-layer framework (chapter 4) · Applies the framework to credit scoring.
Data asset layer (chapter 4) · The data asset layer in the credit scoring framework.

Section by section

Introduction

The introduction frames fraud as a digital, evolving threat and AI as both a defense and a tool for fraudsters.

Shift to digital fraud: Explain how financial theft has shifted from physical cash to exploiting digital systems, and why the growth of online transactions creates more opportunities for fraud.
Why it matters: This shift explains why fraud detection must address activity across digital financial systems.
AI as a double-edged sword: Explain why AI complicates fraud defense: institutions can use it to detect and prevent fraud, while adversaries can use it to make fraud more convincing or operate at greater scale.
Why it matters: A useful fraud defense has to account for adversaries who can improve their methods too.

Key terms: digital transactions, e-commerce, digital payments, online banking, double-edged sword of AI, phishing emails, synthetic identities, cat-and-mouse game

7.1 The criticality of fraud detection and prevention

The section explains why digital growth, evolving fraud, regulation, and customer expectations make effective fraud detection a strategic necessity for financial institutions.

Fraud prevention as strategic imperative: Explain why fraud detection matters beyond preventing direct financial losses, and use the scale of prevention investment to illustrate its urgency.
Why it matters: It frames fraud defenses as part of maintaining a viable, trusted financial service.
Expanding digital fraud exposure: Describe how digital transaction growth and increasingly sophisticated attackers make fraud harder to prevent, including the role of GenAI.
Why it matters: It explains why defenses must keep pace with both transaction volume and changing attack methods.
Customer security friction paradox: Explain the tension between protecting customers from fraud and keeping legitimate transactions convenient.
Why it matters: A defense that blocks fraud but drives away legitimate customers can still fail the business.
Regulatory and systemic fraud pressures: Describe how compliance demands and interconnected financial systems add to the need for effective fraud prevention.
Why it matters: Fraud controls protect institutions against obligations and risks that extend beyond a single transaction.

Key terms: fraud detection and prevention, customer trust, operational resilience, digital transactions, attack surface, GenAI, social engineering, customer expectations, friction paradox, security friction

7.2 Demystifying financial fraud: Scope and typologies

The section defines the book’s scope for financial fraud and organizes six kinds of fraud by their attack vectors to support targeted detection.

Financial fraud scope: Distinguish the broad legal meaning of fraud from the book’s narrower focus on illicit activity that exploits BFSI financial systems, products, or transactions to misappropriate funds or assets.
Why it matters: This boundary keeps the discussion focused on fraud that financial institutions need to detect.
Payment identity and application vectors: Distinguish transaction-based, identity-based, and application-based fraud by what is abused and when, using an example of each.
Why it matters: Identifying what a scheme attacks is a first step toward choosing a relevant detection strategy.
Insider market and human vectors: Distinguish internal or employee fraud, market manipulation fraud, and social engineering-based fraud by whom or what each targets, using an example of each.
Why it matters: These distinctions prevent an insider theft, a market deception, and a phishing attempt from being treated as the same attack.
Attack vector typology purpose: Explain why the section groups fraud by how it is perpetrated rather than presenting only a list of scams.
Why it matters: A useful classification helps connect the kind of attack to the detection problem without relying on a fixed catalog of scams.

Key terms: fraud, financial fraud, BFSI, transaction-based fraud, identity-based fraud, Application-based fraud, Account Takeover (ATO), Synthetic ID Fraud, Internal/employee fraud, Market manipulation fraud

7.3 The distinctive nature of financial fraud

The section explains why financial fraud is a distinctive AI challenge and why anomaly detection is useful but insufficient on its own.

Adversarial behavior and drift: Explain why fraud detection is an active arms race and how fraudsters’ adaptation can make a previously effective detector obsolete.
Why it matters: A fraud defense cannot assume that yesterday’s suspicious patterns will remain reliable.
Fraud class imbalance: Describe how the rarity of fraudulent transactions makes them difficult for traditional ML models to identify.
Why it matters: A detector must contend with fraud being rare even in a large transaction dataset.
Opposing error costs: Distinguish the costs of missing fraud from the costs of blocking a legitimate transaction, and explain why detection involves a tradeoff.
Why it matters: A useful fraud detector must balance protection against disruption to legitimate customers.
Diverse fraud evidence: Describe why fraud detection draws on multiple kinds of data rather than a single straightforward transaction signal.
Why it matters: Relevant evidence about a transaction can take very different forms.
Explainable fraud decisions: Explain why a financial institution must be able to account for a blocked transaction or frozen account.
Why it matters: A decision to stop someone’s financial activity must be defensible to the people it affects and to regulators.
Anomalies versus intentional fraud: Distinguish anomaly detection from fraud detection and explain why anomaly scores alone cannot establish fraud.
Why it matters: An unusual transaction is a clue, not a verdict—and fraud need not be unusual.

Key terms: active arms race, adversarial and organized behavior, rapid concept drift, extreme class imbalance, sparsity problem, majority class, asymmetrical costs of error, false negative, false positive, optimal operating point

7.4 Foundational approaches to fraud detection: The four-layer framework in action

Introduces the four-layer framework as a way to recognize common architectural patterns in financial fraud prevention systems.

Four-layer fraud detection lens: Explain how the four-layer framework provides a lens for understanding the architecture of fraud prevention systems across financial institutions, and name its layers.
Why it matters: The framework gives you a way to organize the fraud detection approaches developed in the following sections.

Key terms: four-layer framework, data asset, modeling, strategy and monitoring, application

7.4.1 A glimpse into the framework: Foundational detection techniques

This section uses simple fraud rules and a chargeback graph investigation to show how foundational detection techniques fit the four-layer framework.

Fraud rules and rule engines: Explain how human-defined if-then conditions flag suspicious activity and how a rule engine combines and applies rules to incoming transactions.
Why it matters: This is the section’s starting point for understanding operational fraud decisions without an AI model.
Rule performance tradeoffs: Describe how analysts test proposed rules on historical transactions and weigh fraud caught against legitimate transactions incorrectly flagged.
Why it matters: A rule that catches fraud can still be unsuitable if it misses most cases or burdens customers and investigators.
Shared links in transaction graphs: Explain how representing users and merchants as nodes connected by transactions can expose a possible shared point of compromise among chargeback cases.
Why it matters: Looking across relationships can reveal a lead that isolated chargeback transactions obscure.
Foundational methods across four layers: Trace how both the rule example and the chargeback investigation move from data and simple analysis through strategic decisions to operational actions.
Why it matters: The framework connects a detection insight to a deliberate, auditable response rather than treating analysis as the final step.

Key terms: rule-based fraud detection, if-then, rule engine, rule sets, detection rate, precision, false positives, insult rate, graph-based pattern discovery, nodes

7.4.2 Data asset layer for AI-powered fraud detection

This section explains what fraud detection needs from the data asset layer: behavioral and relational data, fraud-specific features, and pipelines fast enough to act while fraud is unfolding.

Fraud data versus credit data: Distinguish the data needed for fraud detection from the data emphasized in credit scoring, while noting that both use the same basic data asset layer functions.
Why it matters: The familiar data architecture has to handle a very different kind of evidence when the question is fraud rather than repayment.
Fraud evidence and connections: Describe the main kinds of data gathered for fraud detection and explain why connections across accounts and entities are particularly valuable.
Why it matters: A suspicious transaction may become recognizable only when its session, device, reputation, or connections to other accounts are considered.
Fraud specific feature signals: Describe the kinds of features engineered from fraud data, including rapid activity, deviations from a user’s habits, suspicious links, and session or device trust.
Why it matters: These features make fast-changing and connected fraud clues usable by detection models.
Real time feature pipeline: Trace how streamed fraud events become features and reach models, and explain why the pipeline must respond quickly.
Why it matters: A useful fraud signal must reach a decision-making model while intervention is still possible.

Key terms: data asset layer, behavioral and relational metadata, subsecond telemetry, core transaction data, dynamic customer account and profile information, device and session forensics, enriched external data feeds, network/graph data, hyper-granular velocity features, behavioral anomaly and deviation features

7.4.3 Modeling layer: AI techniques for fraud detection

This section explains how different fraud-detection models address known fraud, novel anomalies, connected fraud rings, and scarce labels, then positions LLMs within a layered fraud-defense system.

Supervised fraud pattern detection: Explain how supervised models learn from labeled fraud and legitimate cases, why they work well for known patterns, and how sparse, imbalanced labels and new tactics limit them.
Why it matters: This explains both the value of historical fraud labels and why a strong classifier cannot be the entire defense.
Unsupervised anomaly detection tradeoffs: Describe how unsupervised methods flag departures from predominantly legitimate behavior and why their ability to surface new fraud comes with false positives and a changing baseline.
Why it matters: Anomaly alerts can reveal emerging attacks, but investigators must distinguish them from unusual legitimate activity.
Relational fraud graph analysis: Explain how graph models represent entities and relationships to expose coordinated fraud that individual-transaction models may miss, and identify their practical complexity.
Why it matters: The connections between otherwise ordinary-looking accounts can be the strongest evidence of organized fraud.
Scarce fraud label remedies: Distinguish using a small labeled set with a large unlabeled pool from generating synthetic fraud examples, including the benefits and risks of artificial data.
Why it matters: These approaches can make limited fraud labels more useful without pretending generated examples are automatically trustworthy.
LLM investigation and grounding: Describe the fraud-management tasks suited to LLMs and how a managed RAG pipeline grounds their case summaries and analysis in controlled sources.
Why it matters: LLMs are most valuable here as evidence-supported aids to threat discovery and investigators, not replacements for every scoring decision.
Layered real-time fraud defense: Justify separating fast transaction blocking from asynchronous LLM analysis, and describe how cost, privacy, and audit requirements influence LLM deployment.
Why it matters: Keeping deep LLM analysis off the immediate authorization path preserves speed while reserving its cost for cases where it adds value.

Key terms: supervised learning, labeled fraud data, class imbalance, cost-sensitive learning, unsupervised learning, anomaly detection, normal behavior, false positive, graph-based analytics, nodes

7.4.4 Strategy and monitoring layer: From scores to actionable intelligence

This section explains how the strategy and monitoring layer turns fraud-model outputs into decisions, human investigations, and continuously governed defenses.

Risk tiers and actions: Explain how the strategy and monitoring layer converts model outputs into actions by setting score thresholds for low-, medium-, and high-risk cases, while accounting for institutional constraints.
Why it matters: Model predictions protect the institution only when they lead to appropriate actions.
Threshold recalibration: Explain why fraud-score thresholds are monitored and adjusted rather than fixed permanently.
Why it matters: A threshold that once balanced protection and customer friction may cease to do so.
Policy rules and lists: Describe how explicit policies and maintained watchlists supplement score-based fraud decisions.
Why it matters: Scores alone cannot enforce every institutional policy or use every piece of known intelligence.
Champion challenger testing: Describe how institutions test a proposed fraud strategy against the current strategy before promoting it.
Why it matters: Testing changes against live outcomes helps avoid improving detection at an unacceptable customer or operational cost.
Analyst review feedback loop: Trace how alerts reach human analysts and how investigation outcomes improve fraud defenses.
Why it matters: Analyst time is limited, and their findings can make subsequent decisions better.
Performance monitoring and governance: Explain what the team monitors and why ongoing validation and governance are needed.
Why it matters: A fraud defense must remain useful and responsible as both attacks and incoming data change.

Key terms: strategy and monitoring layer, risk tiers, actionable thresholds, step-up authentication, thresholds, false positives, fraud rates, business rules, policy overlays, blacklists

7.4.5 Application layer: Deploying and operationalizing fraud defenses

The section explains how the application layer turns fraud scores and strategies into enforced decisions, account protections, analyst workflows, and customer interactions.

Operational role and integration: Explain how the application layer turns upstream fraud intelligence into actions involving customers, analysts, and banking systems, and why integration matters.
Why it matters: A fraud score protects anyone only when it reaches the systems and people that can act on it.
Real-time transaction decisions: Trace how an initiated payment can lead to an approval, decline, or authentication challenge.
Why it matters: Immediate payments require fraud assessments to become enforceable decisions quickly.
Account protection and alerts: Describe automated responses to signs of account takeover and how the application layer communicates suspicious activity to customers.
Why it matters: Fraud defenses must protect accounts while giving legitimate customers safe, useful guidance.
Fraud analyst case management: Explain what a fraud analyst’s case management system presents and how it supports investigation and audit.
Why it matters: Some fraud cases need human review backed by accessible evidence and a record of what was done.
E-commerce chargeback support: Describe the two ways chargeback risk scores can support e-commerce decisions and disputes.
Why it matters: The application layer can support both prevention before fulfillment and the handling of later disputes.

Key terms: application layer, APIs, standardized data protocols, real-time transaction decisioning, fraud scoring, strategic thresholds, step-up authentication, proactive account monitoring, ATO, automated alerting

7.4.6 Synthesizing the layers: A blueprint for operational resilience

This section explains why a decoupled four-layer fraud architecture helps institutions adapt to changing threats without interrupting payment authorization.

Modular fraud system resilience: Explain why separating the data asset, modeling, strategy and monitoring, and application layers makes a fraud system more resilient than a monolithic design. Describe how one component can change without disrupting subsecond payment authorization.
Why it matters: Fraud defenses need to evolve without slowing or breaking live payments.

Key terms: four-layer architecture, operational resilience, structural agility, monolithic systems, subsecond payment authorization flow

7.5 Summary

Where this is used later

Practical realities: The ongoing vigilance in supervised fraud modeling (chapter 8) · Motivates ongoing vigilance in supervised fraud modeling.
(chapter intro) (chapter 8) · Provides earlier context on the distinctive nature of financial fraud.
The cornerstone: Acquiring, managing, and wrestling with labeled data (chapter 8) · Provides context for the difficulties of working with labeled fraud data.
Strategic insight: Analyzing performance at top-N% prediction thresholds (chapter 8) · Informs analysis of predictions at different top-N% thresholds.
Core components: Nodes, edges, and properties (chapter 9) · Develops the graph relationships named as one kind of fraud evidence.
Autoencoders: Learning normality to detect deviance (chapter 8) · Provides the distinction needed when using an autoencoder to detect deviations from normality.
Evaluating the tuned model: Comprehensive metrics on the test set (chapter 8) · Fraud model evaluation likewise requires attention to what is caught and what is incorrectly flagged.
Beyond binary labels: Achieving contextual understanding in fraud classification (chapter 8) · Assumes the data asset layer’s place in fraud detection.

Concepts explained in this chapter

fraud detectionThe process of checking activity or applicants for indicators of fraud.
class imbalanceA condition in which a classification dataset has substantially different numbers of observations in its classes.
defense in depthDefense in depth layers distinct protective controls so that one control can compensate for another's failure or limitations.
graph analyticsA graph-based approach represents entities and their relationships as a graph and analyzes that structure to identify patterns or make predictions.
large language modelA large language model is a language model with many learned parameters, typically trained on extensive text data to process or generate language.
semi-supervised learningSemi-supervised learning uses a limited amount of labeled data together with a larger pool of unlabeled data.
supervised learningSupervised learning trains models on labeled examples to predict outputs for new inputs.
unsupervised learningUnsupervised learning is a machine-learning approach that identifies structure or patterns in data without relying on labeled outcomes.
graphA structure representing entities as nodes and relationships between them as edges.
graph-based pattern discoveryAn analytical technique that represents entities and their relationships as a graph to reveal patterns such as shared connections.
precisionPrecision is the proportion of positive classifications made by a system that are correct.
rule-based systemA rules-based system is a computational system whose behavior is determined by explicitly specified if-then rules.
rules engineA software component or system that manages and executes predefined rules against data or events.
fraudFraud is intentional deception intended to secure an unfair or unlawful gain or cause another party a loss.
concept driftA change over time in the relationship between input data and the target a model is intended to predict.
feature engineeringFeature engineering is the process of constructing informative input variables from raw data for statistical or machine-learning models.
strategy and monitoring layerThe decision-system layer that interprets model signals against business objectives, policies, regulations, and monitored outcomes.
champion-challenger frameworkChampion-challenger testing compares an incumbent model or decision strategy with one or more candidates to evaluate potential improvements or replacements.
thresholdA cutoff value used to divide observations or trigger an action based on a measured value.
independent variableA measurable input variable or derived attribute supplied to a statistical or machine-learning model.
pipelineAn ordered sequence of data-processing or modeling steps that transforms inputs into outputs.
service layerThe architectural layer that integrates analytical outputs into user-facing or operational financial workflows.
anomaly detectionThe identification of observations or patterns that substantially differ from expected behavior, including without labeled training examples.
interpretabilityThe degree to which people can understand the basis of a model’s predictions or decisions.
application fraudApplication fraud is fraud committed by providing deceptive information when applying for a financial product or service.
identity theftIdentity fraud is fraud involving the deceptive use or impersonation of a person's identity.
internal fraudInternal fraud is fraud committed by people within an organization, such as employees misappropriating funds or stealing customer data.
market manipulation fraudFraudulent activity intended to deceive market participants by improperly influencing financial-market prices or trading.
four-layer architectureA book-specific architecture framework that organizes a financial AI system into data asset, modeling, application, and strategy and monitoring layers.
artificial intelligenceArtificial intelligence is the capability of computational systems to perform tasks that typically require human intelligence.
false-positive rateThe proportion of actual negative cases that a classifier incorrectly identifies as positive.
LLMOpsLLMOps is the set of practices for developing, deploying, monitoring, and maintaining systems based on large language models.
centralityThe value assigned to a node by a centrality measure.
community detectionThe identification of groups of graph nodes that are more densely connected to one another than to the rest of the graph.
cost-sensitive learningCost-sensitive learning incorporates differing costs of prediction errors into model training or decision-making.
embedding modelAn embedding model is a model that maps content, such as text, to numerical vector representations that encode semantic relationships.
graph neural networkA neural network that learns from graph-structured data by aggregating information across connected nodes.
prompt engineeringPrompt engineering is the deliberate design and refinement of instructions supplied to generative models to elicit desired outputs.
resamplingChanging a dataset's composition by sampling its observations, often to address class imbalance.
retrieval-augmented generationA method that retrieves relevant external information and supplies it to a generative model when producing an output.

These guides are generated from the text of Manning books by Manning's knowledge graph and link to the sections they draw on. The full books are available on liveBook and at manning.com.