Calibration, Validation and Sensitivity
MODEL LAB · LABORATORY 14
Calibration, Validation
& Sensitivity.
Turn a solvable model into a credible scientific instrument. Reproduce the benchmark, evaluate structural behaviour, vary uncertain parameters and closures, and report how strongly policy conclusions depend on modelling choices.
THREE DISTINCT TESTS
Solve correctly.
Behave credibly.
Calibration, validation and sensitivity answer different questions. None substitutes for the others, and all must be documented before a policy result can be trusted.
Calibration
z(θ, z̄) = z̄
Can the model reproduce the benchmark SAM exactly with no policy shock?
Validation
Model pattern ≈ evidence
Does the model reproduce historical, structural or stylized empirical behaviour?
Sensitivity
R = f(θ, closure, shock)
Do qualitative and quantitative conclusions survive defensible alternative assumptions?
THE CREDIBILITY WORKFLOW
Audit the model
before the headline.
Replicate
Confirm benchmark residuals and quantities within tolerance.
Diagnose
Inspect identities, market residuals, signs and scale.
Validate
Compare model mechanisms with evidence and theory.
Stress test
Vary elasticities, closures, shocks and data assumptions.
Document
Preserve data, code, parameters and scenario definitions.
MODEL CREDIBILITY AUDITOR
Stress-test the
policy conclusion.
Change benchmark error, elasticity uncertainty and closure sensitivity. Complete the documentation checks to calculate a decision-readiness score.
CREDIBILITY VERDICTThe policy direction is robust, but external validation remains incomplete. Treat the magnitude as conditional.
RESEARCH-GRADE STANDARDS
Credibility is built
through transparency.
Traceable data
Document classification, harmonization, balancing and every transformation from source to SAM.
Defensible parameters
Use empirical estimates, meta-analysis or transparent assumptions with plausible ranges.
Multiple closures
Show whether the main result depends on labour, fiscal, investment or external adjustment rules.
Reproducible scenarios
Preserve source code, solver settings, shocks, baselines and result-processing procedures.
SCIENTIFIC CHECKPOINT
What does benchmark
replication establish?
Which conclusion is justified by exact benchmark replication?
Calibration consistency is necessary but not sufficient for empirical credibility.
