Simulate & explore
Impulse responses for any shock the model declares, with parameter overrides for quick "what if" experiments that never touch the saved model.
Macroeconomic modeling, automated and verified
Describe an economy.
Get a solved DSGE model.
GenXEcon turns a plain-language scenario into a fully solved, verified DSGE model. A multi-agent team designs and writes the equations — then a symbolic compiler proves they actually solve, before you ever see a chart.
The LLM proposes, verified mathematics disposes. Nothing about a model's dynamics is templated or faked — every number traces back to an equation the model actually contains.
- 279
- tests, no network required
- 64
- API endpoints
- 1e-6
- agreement with analytical solutions
- 5
- verified reference models
How a model gets made
Four specialists, one self-repairing pipeline
Seven role-specialised agents collaborate on every model. Failures are fed back for self-repair with escalating detail — a model reaches you only after both the mathematics and the economics check out.
-
01
Theory economist
Designs the model from your description: variables, transmission channels, and testable sign predictions grounded in standard theory — "a positive productivity shock should raise output."
-
02
Computational economist
Writes it as a solvable equation system in Dynare-style timing, seeded from a verified reference model (RBC, New Keynesian, small-open-economy, oil-importing hybrid) that it extends rather than rewrites from scratch.
-
03
Symbolic compiler
Parses the equations symbolically, solves the nonlinear steady state, log-linearizes, and solves the rational-expectations system via Klein's (2000) QZ method — checking Blanchard–Kahn determinacy along the way.
-
04
Economics reviewer
Solves the compiled model, computes impulse responses, and checks them against the theorist's sign predictions and textbook results — the Taylor principle, comovement, and more.
If a model can't be made to work, the app says so and suggests a rephrasing. It never falls back to a canned template.
The platform
From a sentence to a defensible result
Everything downstream of the solved model is computed exactly from that solution — not approximated, not simulated when an analytical answer exists.
Business-cycle properties
Exact population moments — volatility, persistence, comovement with output — from the model's Lyapunov solution. No simulation noise.
Policy analysis
Compare Taylor-rule coefficients by their exact implied volatility of inflation, output and the policy rate. Unstable and indeterminate rules are flagged, never silently accepted.
Welfare cost of fluctuations
The standard Lucas (1987) / Otrok (2001) consumption-equivalent welfare cost of business-cycle volatility, built on exact population moments.
Econometrics toolbox
16 methods over any FRED, World Bank or custom series: HP, Baxter–King and Hamilton filters, robust regression, unit-root tests, VAR, structural VAR and cointegration.
Structural estimation
Joint maximum likelihood over any set of parameters, or full Bayesian posteriors by Metropolis–Hastings — each reported with identification diagnostics.
Model-vs-data validation
Confront the model with reality three ways: business-cycle moments, impulse responses against an identified SVAR, and out-of-sample forecast accuracy with Diebold–Mariano tests.
Publication-ready reports
Equations rendered as real mathematics, every chart with a written explanation, business-cycle properties, policy and welfare analysis — in-app or as a PDF.
Standalone Python export
Every model also ships as self-contained Python that solves itself and plots its own impulse responses — with no dependency on the running app.
Depth where it counts
Three layers, one coherent toolchain
1 Data, without a model
Reduced-form econometrics on observed series, independent of any DSGE structure.
- Trend/cycle decomposition — HP, Baxter–King, Hamilton
- OLS with Newey–West HAC and White robust errors
- ADF, KPSS and Phillips–Perron, reconciled into one verdict
- VAR with Granger causality, variance decomposition, forecasts
- SVAR: recursive, long-run (Blanchard–Quah), sign restrictions
- Cointegration: Engle–Granger, Johansen, VECM
2 Fitting the model to data
Estimate a model's own parameters against real observations — and find out whether the data can identify them at all.
- Joint MLE over structural parameters and shock volatilities
- Bayesian posteriors with explicit priors, by adaptive MCMC
- HPD intervals, split-R̂, per-chain Geweke, effective sample size
- Identification diagnostics on every run
- Kalman-filter likelihood from the model's own state space
3 Does it look like reality?
Three comparisons, in increasing order of strictness — because matching moments is easy and forecasting is not.
- Moments: volatilities, persistence, comovement
- IRFs against an identified SVAR, normalised and scored
- Forecasts against a VAR and a random walk
- Diebold–Mariano tests of whether a difference is real
Honest by design
We tell you what the number can't do
Every result reports the assumption it rests on. A Cholesky ordering is required rather than defaulted. Sign restrictions are labelled as set identification. Estimates arrive with identification diagnostics, because an unidentified parameter still produces a confident-looking number that is really just the prior.
"Convergence diagnostics can reveal failure but never prove success."
MCMC estimation
"A mismatch means the model is wrong or the ordering is wrong — the comparison cannot distinguish them."
IRF matching
"Granger causality measures predictive precedence, not structural causation."
VAR analysis
Verified against known solutions
The compiler reproduces analytical RBC and New Keynesian solutions to 1e-6. Estimators are tested by recovering parameters from data they generated themselves; the MCMC sampler is additionally checked against a closed-form Gaussian posterior.
Limitations stated, not buried
The documentation lists what the platform cannot do — first-order solutions only, no risk-adjusted steady state, recursive SVARs being structurally incompatible with dense DSGE impact matrices — in plain language, up front.
Pricing
Start free. Scale when the research does.
Every plan includes the full solver, the complete econometrics toolbox, and standalone Python export. Nothing about the mathematics is gated.
Researcher
Free
For students, individuals and replication work.
- Unlimited local models
- Full solver, econometrics and estimation
- FRED & World Bank data integration
- Standalone Python export
- Bring your own LLM API key
- Community support
Professional
$49/month
For applied economists shipping real analysis.
- Everything in Researcher
- Hosted workspace — no local install
- Managed model generation, no API key needed
- Publication-ready PDF reports
- Bayesian estimation with priority compute
- Email support, next business day
Institution
Custom
For central banks, faculties and research teams.
- Everything in Professional
- Shared team model library
- SSO and role-based access
- On-premise or private-cloud deployment
- Model review and onboarding workshops
- Dedicated support engineer
Academic and non-profit discounts available. Open-source under the MIT licence — self-host the whole platform at no cost, forever.
Questions
Frequently asked
What is a DSGE model?
A dynamic stochastic general equilibrium model describes an entire economy as a system of equations derived from optimising households and firms, driven by random shocks. Solving one means finding how every variable responds over time to each shock — which requires computing a steady state, linearizing around it, and solving the resulting rational-expectations system.
Is this just an LLM writing equations?
No. The LLM proposes a model; a symbolic compiler then has to prove it solves — steady state, log-linearization, Blanchard–Kahn determinacy. If it doesn't compile, or its impulse responses contradict the theory it claimed, the model is sent back for repair. Nothing reaches you on the LLM's word alone.
Which models can it build?
Anything expressible as a system of equations in Dynare-style timing. Generation is seeded from five verified reference models — RBC, New Keynesian, New Keynesian with capital, small open economy, and an oil-importing hybrid — which the agents extend rather than rewrite.
Do I need an API key?
For model generation, yes — an OpenAI key on the free plan, or use our managed generation on Professional. World Bank data needs no key at all. FRED needs a free key, and the app tells you exactly how to get one if it's missing.
Can I trust the numbers?
The solver is verified against known analytical solutions to 1e-6, and the estimators are tested by recovering parameters from data they generated. More importantly, every result states the assumption it depends on — and the documentation lists what the platform cannot do, in plain language.
Is my work locked in?
No. Every model exports as self-contained Python that solves itself and plots its own impulse responses, with no dependency on the running app. Models are stored as plain JSON. The platform is MIT licensed and can be self-hosted in full.
Does it forecast better than a VAR?
Often not — and the platform measures exactly that rather than hiding it. Structural models describe mechanisms; forecasting rewards capturing persistence. The forecast comparison runs the model against a VAR and a random walk and reports Diebold–Mariano tests of whether the difference is real.
Put a model in front of the data today
Free to start, MIT licensed, and honest about what it can't do.