AI Policy Modelling for Central Banks Training Course

Participants can translate a policy question into a measurable modelling specification.

• Policy questions converted into targets, horizons, decision thresholds, and model outputs
• Nowcasting, forecasting, classification, and anomaly detection matched to policy decisions
• Target-variable definitions tested for revisions, publication lags, and measurement changes
• Prediction objectives separated from causal inference and structural policy evaluation
• Benchmark models selected before machine-learning alternatives are developed
• Success criteria defined using accuracy, timeliness, interpretability, and operational feasibility

Participants can design a data structure for AI-supported policy analysis.

• National accounts, price indices, labour data, and monetary aggregates catalogued
• Market yields, exchange rates, credit spreads, and commodity prices aligned
• Bank supervisory returns incorporated with confidentiality and access restrictions
• Release calendars used to distinguish available data from revised observations
• Data dictionaries record definitions, transformations, units, ownership, and refresh schedules
• Metadata controls document source quality, licensing, lineage, and revision policies

Participants can prepare macroeconomic datasets for reproducible model development.

• Missing observations treated using interpolation, imputation, and explicit availability flags
• Seasonal adjustment assessed using X-13ARIMA-SEATS outputs and revision patterns
• Outliers investigated before winsorisation, transformation, or exclusion decisions
• Stationarity checked using unit-root tests, differencing, and growth-rate transformations
• Feature tables constructed with publication lags and real-time information sets
• Train, validation, and test samples separated by policy-relevant time periods

Participants can create economically meaningful features for forecasting and monitoring models.

• Lagged variables constructed to reflect transmission delays and reporting schedules
• Rolling averages, growth rates, gaps, and acceleration measures calculated consistently
• Yield-curve slopes and credit spreads derived as financial-conditions indicators
• Inflation expectations extracted from surveys, market instruments, and text indicators
• Interaction terms tested for nonlinear responses during tightening and easing cycles
• Feature selection constrained by economic plausibility, availability, and interpretability

Participants can establish econometric benchmarks for machine-learning model comparison.

• Autoregressive models estimated for inflation, output, and unemployment forecasting
• Phillips curve specifications tested with slack, expectations, import prices, and lags
• Taylor-rule estimates used to frame historical policy-rate responses
• Vector autoregressions applied to output, prices, rates, and exchange rates
• Dynamic factor models used to summarise large macroeconomic datasets
• Forecast benchmarks documented before machine-learning performance is assessed

Participants can select machine-learning methods appropriate to economic policy questions.

• Supervised learning framed through labelled outcomes and prediction-loss functions
• Unsupervised learning applied to regime identification and indicator clustering
• Bias-variance trade-offs related to sample size, noise, and policy stability
• Regularisation methods compared using LASSO, Ridge, and Elastic Net penalties
• Cross-validation adapted for ordered economic time-series observations
• Hyperparameters controlled through documented search ranges and validation procedures

Participants can build a machine-learning nowcast for near-term inflation assessment.

• Mixed-frequency indicators aligned with monthly inflation publication schedules
• Elastic Net models estimated using prices, wages, exchange rates, and surveys
• Random forests capture nonlinear relationships among inflation drivers
• Gradient boosting models tested against autoregressive inflation benchmarks
• Nowcast updates linked to data releases and forecast revisions
• Forecast errors decomposed by food, energy, core, and administered-price components

Participants can develop forecasts for economic activity using structured machine-learning methods.

• Monthly activity indicators combined with quarterly GDP measurement frameworks
• Labour-market features include vacancies, earnings, hours, and participation rates
• Tree-based models estimate nonlinear demand responses to financial conditions
• Gradient boosting compared with dynamic factor model forecasts
• Recession classification evaluated using precision, recall, and calibration metrics
• Forecast horizons matched to monetary-policy report and committee schedules

Participants can develop AI-supported indicators for financial conditions and systemic risk.

• Financial conditions indices constructed from rates, spreads, equities, and currencies
• Banking-sector indicators include capital, liquidity, asset quality, and profitability
• Anomaly detection identifies unusual movements in supervisory and market data
• Clustering methods distinguish stress episodes, credit cycles, and market regimes
• Early-warning models tested against historical banking and liquidity events
• False alarms assessed against supervisory escalation and intervention costs

Participants can extract structured policy indicators from central-bank and market text.

• Text corpora prepared from statements, minutes, speeches, news, and surveys
• Tokenisation, stemming, and stop-word treatment documented for reproducibility
• Sentiment scores derived from central-bank communications and financial news
• Topic modelling identifies recurring concerns in inflation and financial-stability narratives
• Text embeddings compared for semantic similarity across policy communications
• Language-model outputs checked for factual reliability, bias, and source traceability

Participants can evaluate policy models using disciplined out-of-sample testing.

• Rolling-origin evaluation mirrors real-time forecasting and data-release conditions
• Root mean squared error compared across competing forecast models
• Mean absolute error used where large misses require limited influence
• Diebold-Mariano tests assess statistically meaningful forecast performance differences
• Directional accuracy measured for inflation turning points and growth slowdowns
• Backtest results segmented by shocks, regimes, forecast horizons, and revisions

Participants can interpret AI model outputs without overstating causal claims.

• SHAP values attribute individual forecasts to contributing model features
• Partial dependence plots show average feature effects under stated assumptions
• Individual conditional expectation plots reveal heterogeneous feature-response patterns
• Feature importance checked against multicollinearity, instability, and economic theory
• Causal claims withheld when models only demonstrate predictive association
• Economists challenge model narratives using institutional knowledge and counterfactual reasoning

Participants can integrate AI forecasts into policy scenarios and decision narratives.

• Baseline projections linked to published assumptions for global and domestic conditions
• Alternative scenarios specify shocks to energy, exchange rates, demand, and credit
• Machine-learning forecasts stress-tested under plausible but unfamiliar input combinations
• Scenario narratives distinguish conditional projections from policy recommendations
• Fan charts communicate forecast uncertainty across policy-relevant horizons
• Policy implications assessed against inflation targets, output risks, and financial stability

Participants can prepare a validation framework for AI policy models.

• Conceptual soundness reviews test objectives, assumptions, features, and model choice
• Independent validation examines code, data lineage, outputs, and documentation
• Sensitivity analysis tests exposure to revisions, outliers, and parameter changes
• Data leakage checks prevent future information entering historical model estimates
• Performance thresholds trigger recalibration, escalation, redevelopment, or model retirement
• Validation reports record limitations, compensating controls, findings, and approvals

Participants can define governance controls for responsible AI use in central banking.

• Model inventories classify policy models by materiality, complexity, and decision use
• Approval authorities assigned across business owners, risk teams, and committees
• Bias assessments examine distributional impacts across regions, households, and firms
• Confidential data controls cover access, retention, encryption, and audit logging
• Third-party model risks assessed through documentation, testing, and contractual controls
• Human review requirements defined for policy recommendations and public communications

Participants can plan controlled deployment and monitoring of policy models.

• Reproducible workflows use version control, parameter files, and documented dependencies
• Model runs scheduled around release calendars and policy-committee deadlines
• Monitoring dashboards track accuracy, data quality, drift, and availability
• Champion-challenger models support periodic comparison against established benchmarks
• Exception procedures address missing data, model failures, and unexpected forecasts
• Policy brief templates link model outputs, uncertainty, risks, and recommended follow-up

Participants can build, review and present a complete AI-supported policy model for a defined central-bank decision.

• Policy problem framed with target variable, horizon, users, and decision context
• Real-time dataset prepared with documented sources, transformations, and release lags
• Benchmark econometric and machine-learning models compared through rolling backtests
• Model explanations assessed using SHAP values, diagnostics, and economic reasoning
• Governance pack prepared with validation findings, limitations, monitoring, and approvals
• Policy committee briefing presents scenarios, uncertainty, recommendations, and challenge responses