AI Fraud Detection Training Course

Participants can prioritise fraud problems suitable for AI-assisted detection.

• Payment fraud typologies: card-not-present, authorised push payment and mule activity
• Account takeover indicators across authentication, device and transaction events
• Claims fraud patterns involving duplication, exaggeration and collusive networks
• Procurement fraud signals within suppliers, invoices, approvals and payments
• Risk-based use-case scoring using loss, volume, detectability and data readiness
• Use-case hypotheses written with measurable fraud and operational objectives

Participants can define fit-for-purpose data requirements for fraud detection.

• Transaction, customer, device, merchant and case-management data mapped
• Entity resolution matches customers, accounts, devices and counterparties
• Data lineage documents source systems, transformations and ownership
• Missing values, duplicates and timestamp errors assessed systematically
• Label quality reviewed using confirmed fraud, chargebacks and case outcomes
• Training, validation and test datasets separated by time period

Participants can specify interpretable features that capture fraud behaviour.

• Velocity features measure transaction counts, amounts and failed attempts
• Behavioural baselines compare current activity against customer history
• Geographic distance features identify improbable location and travel patterns
• Merchant and counterparty risk features incorporate prior fraud outcomes
• Time-based features capture unusual hours, sequences and transaction bursts
• Feature dictionaries record calculation logic, purpose and data provenance

Participants can assess supervised classification models for labelled fraud data.

• Binary classification predicts fraudulent versus legitimate transaction outcomes
• Logistic regression provides transparent baseline fraud risk scores
• Decision trees identify conditional patterns within fraud-labelled populations
• Random forests reduce variance across multiple decision-tree predictions
• Gradient boosting ranks complex interactions among fraud risk indicators
• Class imbalance addressed through weighting, sampling and threshold selection

Participants can apply anomaly detection where confirmed fraud labels are limited.

• Unsupervised detection identifies observations distant from normal behaviour
• Isolation Forest isolates rare transaction patterns through random partitions
• Local Outlier Factor compares density around neighbouring observations
• Clustering segments customers before identifying unusual within-segment behaviour
• Anomaly scores combined with rules to improve investigator relevance
• Novelty detection separates emerging patterns from known fraud typologies

Participants can identify network structures associated with organised fraud.

• Graph models represent entities as nodes and relationships as edges
• Connected-component analysis identifies linked accounts, devices and beneficiaries
• Centrality measures identify influential intermediaries within suspicious networks
• Community detection reveals clusters of coordinated transactional behaviour
• Mule-account networks traced through shared devices, addresses and transfers
• Network features incorporated into transaction-level fraud risk scores

Participants can identify practical uses of text analysis in fraud controls.

• Text classification routes case narratives, complaints and investigation notes
• Named-entity recognition extracts people, organisations, accounts and locations
• Keyword dictionaries identify scam narratives and emerging fraud terminology
• Document similarity detects repeated invoices, claims narratives and evidence
• Sentiment signals interpreted cautiously within customer-contact records
• Human review validates text-model outputs before investigative action

Participants can combine analytical methods into an operational detection strategy.

• Decision strategies combine model scores, rules and contextual data
• Risk bands route alerts to automated action or investigator review
• Thresholds set according to fraud loss, workload and customer impact
• Champion-challenger models compare proposed methods against current controls
• Decision matrices specify approve, decline, hold and escalate actions
• Fallback rules maintain controls during model failure or data outages

Participants can design alert workflows that support timely, evidence-based decisions.

• Alert queues prioritised by predicted loss, confidence and urgency
• Case-management fields capture evidence, rationale and final disposition
• Investigation playbooks define checks for common fraud scenarios
• Service-level targets distinguish real-time intervention from retrospective review
• Feedback loops return investigator outcomes to model-development teams
• Quality assurance samples closed cases for consistency and accuracy

Participants can evaluate fraud model performance using operational and financial measures.

• Confusion matrices classify true positives, false positives and missed fraud
• Precision measures the proportion of alerts confirmed as fraud
• Recall measures confirmed fraud captured from the total fraud population
• ROC curves compare discrimination performance across possible score thresholds
• Precision-recall curves assess performance under severe class imbalance
• Expected-value analysis balances prevented losses, review costs and customer friction

Participants can plan validation and calibration for fraud detection models.

• Out-of-time testing assesses performance against later transaction periods
• Back-testing compares predicted risks with confirmed fraud outcomes
• Sensitivity analysis tests threshold changes under varying fraud rates
• Stress scenarios test performance during fraud attacks and data disruption
• Calibration plots compare predicted probabilities with observed fraud frequencies
• Independent validation challenges assumptions, data choices and performance claims

Participants can produce explanations that support challenge and investigation decisions.

• Global explainability identifies features driving overall model predictions
• Local explanations identify factors behind individual alert scores
• SHAP values quantify each feature's contribution to a prediction
• Reason codes translate model signals into investigator-friendly language
• Counterfactual analysis identifies changes that would alter a decision
• Explanation records retained for challenge, audit and customer complaints

Participants can identify responsible-AI risks in fraud detection decisions.

• Fairness testing compares error rates across relevant customer groups
• Proxy variables assessed for indirect discrimination and unintended bias
• Data minimisation limits personal data to defined fraud purposes
• Purpose limitation governs reuse of customer and behavioural information
• Human oversight requirements defined for consequential customer decisions
• Responsible-AI impact assessments document harms, controls and residual risks

Participants can document governance controls for AI fraud detection models.

• Model inventories record purpose, owner, version and risk rating
• SR 11-7 principles applied to model development and validation
• Approval committees review performance, limitations and control evidence
• Model documentation records assumptions, features, thresholds and dependencies
• Change-control procedures govern retraining, releases and emergency adjustments
• Audit trails preserve inputs, decisions, overrides and reviewer actions

Participants can design monitoring for changing fraud patterns and model performance.

• Data drift detects changes in input distributions and data quality
• Concept drift identifies altered relationships between signals and fraud
• Population Stability Index measures shifts between reference and live data
• Performance dashboards track precision, recall, losses and alert volumes
• Retraining triggers define evidence thresholds, approvals and rollback requirements
• Business continuity plans maintain detection during technology outages

Participants can assess vendor-provided AI fraud detection solutions.

• Vendor due diligence assesses data use, model transparency and security
• Service-level agreements define uptime, response times and support responsibilities
• Contractual audit rights support independent control and performance review
• Model documentation requested for features, testing and known limitations
• Data-processing agreements address retention, access and cross-border transfers
• Exit plans protect continuity if providers fail or contracts end

Participants can plan implementation across fraud operations, technology and governance teams.

• Target operating models assign accountabilities across first and second lines
• RACI matrices define responsibilities for models, alerts and investigations
• Implementation roadmaps sequence data, model, workflow and governance deliverables
• Pilot designs establish scope, success criteria and control boundaries
• Change-management plans prepare investigators for revised alert decisioning
• Management information reports communicate value, risk and operational performance

Participants can diagnose detection failures and define corrective control actions.

• Root-cause analysis separates data, model, process and governance failures
• Post-incident reviews reconstruct missed fraud and delayed intervention decisions
• Attack-pattern analysis identifies adversarial behaviour against detection controls
• Control gaps prioritised using loss exposure, recurrence and remediation effort
• Lessons-learned registers assign owners, deadlines and verification evidence
• Scenario libraries updated with emerging patterns and confirmed case findings

Participants can review an AI fraud detection design against control and operational criteria.

• Detection design reviews test alignment with defined fraud risk scenarios
• Data readiness checklists confirm availability, quality, lineage and permissions
• Model selection rationales compare accuracy, explainability and operational fit
• Alert capacity models test staffing against expected detection volumes
• Governance reviews confirm approvals, validation evidence and accountability
• Implementation decisions document dependencies, residual risks and next actions

Participants can build, review and present a complete AI fraud detection operating design for a defined fraud scenario.

• Fraud scenario defined with losses, actors, behaviours and control objectives
• Data specification maps sources, labels, features, quality and access controls
• Detection strategy combines rules, model scores and network indicators
• Alert workflow assigns thresholds, queues, investigations and escalation paths
• Validation plan defines metrics, testing, monitoring and retraining triggers
• Governance pack documents explainability, fairness, approvals and audit evidence