Home / Methods
Methodology

Built on methods your review committee recognises.

PolicySandbox is not a black box. What runs today is a calibrated semi-structural causal engine with distributed lags, solved in deviation-from-baseline space, with Monte Carlo over coefficient uncertainty and a household microsimulation on top. The wider toolkit below is listed with an honest status against each entry, because the fastest way to lose an economist is to claim a DSGE you do not have.

In the engineImplemented in the engine running at policysandbox.ai today.Calibrated fromNot implemented here. The published estimates this method produces are what this engine's coefficients are calibrated against.RoadmapNot in the product. Named because it is where this is going, not because it is running.

Structural macro models

DSGE

Roadmap

Dynamic stochastic general equilibrium: micro-founded households and firms optimising under rational expectations. Policy shocks propagate with full internal consistency.

Powers: Deep scenario engines

Semi-structural / hybrid

In the engine

MARTIN-class models: econometrically estimated equations with theory-imposed long-run structure. The workhorse of official Australian policy analysis.

Powers: Core simulation engine

Large-scale macroeconometric

Roadmap

Hundreds of estimated behavioural equations with rich sectoral detail for budget-grade costings and forecasts.

Powers: Sector detail modules

CGE

Roadmap

Computable general equilibrium: multi-sector, multi-region resource reallocation. The standard for tax, trade and energy-transition policy.

Powers: Tax & industry levers

Overlapping generations (OLG)

Roadmap

Intergenerational structure for superannuation, pensions, housing and demographic policy questions.

Powers: Retirement & housing levers

Empirical & time-series

SVAR / BVAR / FAVAR

Calibrated from

Identified vector autoregressions estimating how the economy historically responds to policy shocks.

Powers: Elasticity estimation

Local projections

Calibrated from

Jorda-style impulse responses: the modern default for tracing a shock's path over time.

Powers: Transmission calibration

VECM / cointegration

Calibrated from

Long-run equilibrium relationships with short-run error-correction dynamics.

Powers: House price & income anchors

Nowcasting & factor models

Roadmap

Dynamic factor models, MIDAS and bridge equations reading the real-time state of the economy from high-frequency data.

Powers: Live baseline conditions

Micro-founded simulation

Microsimulation

In the engine

Tax-and-transfer engines run over unit-record household data: who wins, who loses, dollars per week.

Powers: Family ledger

Heterogeneous agents (HANK)

Roadmap

Monetary policy with realistic wealth and income distributions: mortgage holders, renters and savers respond differently.

Powers: Mortgage vs renter split

Agent-based models

Roadmap

Bottom-up simulation of adaptive interacting agents with no equilibrium assumption; suited to housing dynamics and contagion.

Powers: Housing market dynamics

Scenario & risk

Stress testing

Roadmap

APRA / FSAP-style adverse scenario propagation through balance sheets and default channels.

Powers: Bank & lender modules

Scenario analysis & fan charts

In the engine

Baseline plus alternative paths with probability bands: uncertainty, communicated honestly.

Powers: Briefing outputs

Monte Carlo simulation

In the engine

Thousands of stochastic runs over shock distributions: outcome distributions, not point estimates.

Powers: Confidence bands

Input-output multipliers

Roadmap

Leontief tables tracing sectoral and employment flow-through of spending shocks.

Powers: Regional & sector impact

Gravity & trade models

Roadmap

Bilateral trade-flow structure for tariff and trade-policy scenarios.

Powers: Trade levers

ML & hybrid forecasting

Roadmap

Gradient boosting, neural components and causal ML disciplined by structural models.

Powers: Forecast accuracy

Policy evaluation

Quasi-experimental methods

Calibrated from

Difference-in-differences, regression discontinuity, synthetic control and event studies: what a policy actually did, feeding back into calibration.

Powers: Model discipline

4 of 19 run in the engine today; 4 are what its coefficients are calibrated against; 11 are roadmap. What exists now is documented in full - 90 registered models over 132 variables and 342 relationships, each with an elasticity, a lag, a confidence class and its limitations written down.

Talk methodology with us.

Enterprise engagements include full model documentation: specification, sources, elasticities and validation. Bring your economists - the hard questions are the useful ones.

Launch the sandbox