Machine Learning for Banking Decision Making Training Course

Participants can assess data fitness and control requirements before model development.

• Core banking, credit bureau, transaction and customer interaction data
• Data lineage from source system through feature store and model
• Completeness, accuracy, timeliness and consistency quality checks
• Missing-value patterns distinguished from operationally meaningful customer behaviour
• Personally identifiable information handled through minimisation and access controls
• Data dictionaries documenting fields, definitions, ownership and permitted usage

Participants can create and document meaningful model features from banking data.

• Debt-service ratio, utilisation rate and delinquency trend features
• Transaction velocity, merchant concentration and unusual pattern indicators
• Rolling windows built for balances, payments and account activity
• Categorical variables encoded without losing operational interpretability
• Outliers treated using capping, transformations and investigation rules
• Feature documentation linking calculation logic to business rationale

Participants can prevent common data and sampling errors that invalidate banking models.

• Target leakage identified through post-decision events and outcome fields
• Time-based splits protecting against future information entering training data
• Reject inference limitations in credit application scorecard development
• Class imbalance handled with weighting, sampling and appropriate metrics
• Selection bias arising from channel, product and customer eligibility rules
• Protected characteristics assessed for direct and proxy discrimination risks

Participants can select and compare classification methods for lending decisions.

• Logistic regression applied to approval, default and arrears prediction
• Decision trees interpreted through explicit customer segmentation rules
• Random forests compared with single trees for predictive stability
• Gradient boosting assessed for performance, complexity and explainability
• Probability of default outputs connected to credit approval policies
• Champion-challenger design supporting controlled model comparison

Participants can design machine learning approaches for fraud and financial crime alerting.

• Fraud labels defined from confirmed cases, investigations and chargebacks
• Anomaly detection used where reliable labelled outcomes remain limited
• Isolation Forest identifies unusual transaction behaviour and account activity
• Network features reveal linked counterparties, mule accounts and collusive patterns
• Alert ranking aligned with investigator capacity and case prioritisation
• Precision at k measures value within constrained investigation queues

Participants can adapt machine learning methods to customer and insurance decisions.

• Churn prediction identifies customers requiring retention treatment or outreach
• Propensity models prioritise suitable offers under conduct risk constraints
• Collections segmentation distinguishes ability-to-pay from willingness-to-pay indicators
• Survival analysis estimates time until arrears, lapse or claim events
• Claims triage ranks cases for straight-through processing or investigation
• Treatment uplift separates intervention impact from natural customer behaviour

Participants can organise model development so results can be reproduced and reviewed.

• Training, validation and test sets assigned using chronological cutoffs
• Cross-validation applied when observations are sufficiently stable and independent
• Hyperparameter tuning controlled through documented search ranges and selection criteria
• Random seeds, code versions and data snapshots recorded
• Pipelines standardise transformations between development and production environments
• Development notebooks converted into reviewable model development documentation

Participants can evaluate model performance using statistical and banking decision measures.

• Confusion matrices translate predictions into approval and intervention outcomes
• ROC-AUC measures ranking performance across possible classification thresholds
• Precision-recall curves assess models with rare fraud and default events
• Kolmogorov-Smirnov statistics compare score distributions between outcome groups
• Calibration plots test whether predicted probabilities match observed outcomes
• Population Stability Index detects material shifts in score distributions

Participants can set model decision thresholds using risk appetite and economic criteria.

• Expected loss combines probability of default, loss given default
• Approval cutoffs reflect risk appetite, capital and profitability constraints
• Fraud thresholds balance prevented losses against customer friction costs
• Queue capacity limits incorporated into investigation and collections prioritisation
• Decision matrices define automatic approval, referral, decline and review actions
• Scenario analysis tests threshold effects under changing portfolio conditions

Participants can explain machine learning decisions to managers, customers and reviewers.

• Global feature importance identifies drivers of overall model behaviour
• SHAP values explain individual predictions using additive feature contributions
• Partial dependence plots reveal average response patterns and nonlinear effects
• Reason codes translated into accurate, actionable customer communications
• Counterfactual explanations identify changes associated with alternative model outcomes
• Explanation limits documented for complex models and correlated features

Participants can test fairness risks and define responsible decision-use controls.

• Disparate impact ratios compare outcomes across relevant customer groups
• Equal opportunity tests examine outcome differences among eligible applicants
• Proxy variables reviewed for indirect discrimination and unfair exclusion
• Vulnerable customer considerations incorporated into decision and escalation rules
• Human override processes controlled through authority, rationale and audit trails
• Conduct risk assessments connect model choices to customer outcomes

Participants can prepare governance evidence for model approval and ongoing oversight.

• SR 11-7 principles applied to model inventory and governance
• Model tiering considers materiality, complexity, customer impact and regulatory exposure
• Independent validation challenges conceptual soundness, data and outcomes
• Model documentation records purpose, methodology, limitations and approved use
• Approval committees assign ownership, conditions and remediation actions
• Exceptions tracked through issue logs, deadlines and accountable owners

Participants can design controlled deployment arrangements for machine learning decision models.

• Batch scoring and real-time scoring compared for banking processes
• Application programming interfaces connect models with workflow and case systems
• Decision rules combine model scores, policy rules and manual referrals
• Fallback procedures protect customers during outages and degraded performance
• User interfaces present scores, explanations and required next actions
• Change management prepares frontline teams for automated decision support

Participants can establish ongoing monitoring and change controls for deployed models.

• Input drift monitored through feature distributions and missing-value rates
• Concept drift identified when outcome relationships change over time
• Performance monitoring tracks discrimination, calibration and decision outcomes
• Fairness metrics reviewed after portfolio, policy or channel changes
• Trigger thresholds define investigation, recalibration and redevelopment actions
• Monitoring dashboards assign reporting frequency, owners and escalation routes

Participants can build, review and present a complete governed machine learning decision case for a banking use case.

• Select a credit, fraud, collections or customer decision problem
• Define target outcome, decision workflow and measurable business value
• Specify data sources, quality checks, features and leakage controls
• Compare candidate models using validation metrics and decision economics
• Set thresholds, human review rules and customer communication requirements
• Document explainability, fairness tests, limitations and model risk controls
• Present implementation, monitoring and approval recommendations to a review panel