red-data-analysis

Quantitative Analysis for RED (red-data-analysis)

When to trigger

• Calibrating or estimating a dynamic model and reporting model-vs-data fit
• Building the numerical experiments that carry a computational paper
• Making the computation reproducible ahead of the RED data/code archive

What RED-quality analysis looks like

RED is method-defined, so "data analysis" usually means disciplining a dynamic model, and for purely theoretical papers it is lighter — focus there on numerical examples and reproducible computation that illustrate the result rather than estimate it.

• Calibration — list every parameter, its value, its source/target, and whether it is calibrated, estimated, or assumed. Calibration targets should be explicit moments, not vibes.
• Moment-matching / estimation — report the targeted moments, the fit (model vs data), and untargeted moments the model also matches (a strong credibility signal). Document the estimator and its assumptions.
• Numerical solution discipline — state the solution method (perturbation, projection, global, EGM, sequence-space), the grid/approximation order, and an accuracy check (e.g., Euler-equation errors).
• Reproducible computation — fix and record random seeds, capture software/OS and package versions, and write a master "run-all" script. RED's readme requires execution order, expected runtime, and seeds, so build that discipline in from the start.

Checklist

• Every parameter has a value and a documented source/target
• Targeted and untargeted moments reported; model-vs-data fit shown
• Solution method, approximation order, and an accuracy check are stated
• Seeds fixed/recorded; environment captured; a run-all script exists

Anti-patterns

• Reporting only the moments the model was calibrated to hit
• An undocumented numerical solution with no accuracy diagnostic
• Leaving seeds and environment uncaptured, so the archive will not reproduce

RED analysis table skeleton

Require a compact audit table before results are written:

Item	RED-ready entry
Parameters	value, source/target, calibrated vs estimated vs assumed
Data moments	targeted moments plus untargeted validation moments
Solver	method, grid/order, convergence rule, accuracy diagnostic
Experiments	baseline, counterfactual, transition path, sensitivity
Archive	run-all command, seeds, software/OS, expected runtime

This table can become a manuscript table, appendix table, or replication README section. If it cannot be filled, the quantitative core is not ready for RED.

Illustrative calibration record (Aiyagari-style economy)

Both the paper and the archive should carry a calibration record in this shape (values illustrative):

Internally calibrated (chosen to hit model-generated targets):
  beta  = 0.957   target: K/Y = 2.9 (annual)                     model hits: 2.90
  phi   = -1.1    target: share with negative net worth = 13.5%  model hits: 13.2%
Externally set (taken from outside estimates):
  sigma = 2.0     CRRA; standard in the incomplete-markets literature (author-year)
  rho_z = 0.91    AR(1) earnings persistence, PSID estimates (author-year)
Accuracy at this parameterization:
  max log10 |Euler error| = -4.6, 200-pt log-spaced asset grid, a_max = 40x mean income

Separating internally calibrated from externally set parameters matters at RED: referees check whether each internal target is actually informative about its parameter, and whether external values are sourced rather than convenient.

Moment-fit objections and the RED-grade response

Objection	Fix
"Targets are not independent of the moment you claim to explain"	re-calibrate with the headline moment excluded; report it as untargeted
"Fit shown only for targeted moments"	add an untargeted panel (wealth Gini, top shares, MPC distribution)
"No accuracy diagnostic at the headline solution"	Euler or den Haan errors at exactly that parameterization
"Grid and discretization undocumented"	state grid sizes, bounds, Rouwenhorst/Tauchen choice, and a grid-doubling stability check

Minimum reproducible run

Before submission, maintain two runs:

• Smoke run: reduced grid or shorter simulation that completes quickly and verifies code paths.
• Publication run: full grid, full simulation length, final seeds, final moments, and final exhibits.

Both runs should write comparable output names and logs. If the smoke run and publication run disagree in qualitative direction, the numerical method needs diagnosis before the paper is ready. Record expected runtime for both because RED's readme requirement makes runtime part of the archive, not an afterthought.

Output format

[Analysis status] ready / missing moments / weak solver diagnostics / unreproducible
[Calibration/estimation] <targets, parameters, estimator>
[Numerical checks] <solver, accuracy, convergence, sensitivity>
[Archive readiness] run-all / seeds / OS / runtime / missing
[Next analysis] <single computation or table to repair>

Supplementary resources

• ../../resources/external_tools.md — Dynare, SSJ, Julia/QuantEcon, FRED
• red-replication-and-data-policy — the archive these outputs feed