Docs calibration howto#296
Open
thorek1 wants to merge 732 commits into
Open
Conversation
… and skip comparisons for higher-order IRFs and variance when applicable
… distinct constants and enhance benchmark reporting for first-order solves Co-authored-by: Copilot <copilot@github.com>
…alculations and update related comments Co-authored-by: Copilot <copilot@github.com>
…rameters for NESSAI
…ditionally include moments
* Add Schur deflation for unit-root covariance computation Models with unit-root eigenvalues (e.g., FRBUS with 23 unit roots) now compute partial covariance matrices via Schur deflation. The approach: 1. When all standard Lyapunov solvers fail, decompose A = U*T*U' via real Schur factorization and reorder so unstable eigenvalues (|λ| ≥ 1 - 1e-8) come first. 2. Solve the Lyapunov equation only for the stable (lower-right) sub-block. 3. Map back to original coordinates; variables with any loading on unstable Schur vectors get NaN (infinite unconditional variance). Changes: - lyapunov.jl: Add solve_lyapunov_schur_deflation() and fallback in dispatch - get_functions.jl: Optimize get_variance_decomposition with pre-computed Schur for per-shock solves; soften @Assert to @warn for covariance failures Verified with FRBUS (89 stationary vars, 339 NaN), RBC, SW07, Guerrieri_Iacoviello. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * Replace LinearAlgebra.schur/ordschur with FastLapackInterface.gees! in Schur deflation Use FastLapackInterface.SchurWs and LAPACK.gees! with eigenvalue selection (FastLapackInterface.ed) to perform Schur decomposition and eigenvalue reordering in a single LAPACK call, replacing the separate schur() + ordschur() sequence (lyapunov.jl) and eigvals() + schur() + ordschur() sequence (get_functions.jl). gees! union return type behind concrete type assertions - Use FastLapackInterface.ed with criterium=(1-tol)^2 to select unstable eigenvalues (exterior of disk) into the top-left Schur block - Read n_unstable from workspace sdim field instead of counting eigenvalues Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * Thread SchurWs workspace through lyapunov_workspace struct Add FastLapackInterface.SchurWs{T} field to lyapunov_workspace, initialized in the lyapunov_workspace struct instead. - Add schur_ws field to lyapunov_workspace struct (structures.jl) - Initialize with SchurWs(zeros(T,1,1)) in Lyapunov_workspace() constructor and get_variance_decomposition - Move ensure_lyapunov_workspace! call before Schur decomposition in get_variance_decomposition so the workspace is available for both paths Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * Add early unit-root detection and fast Lyapunov dispatch Detect unit roots from QME solution eigenvalues during first-order solve, before any covariance computation. This avoids 3 wasted O(n³) Lyapunov solver attempts for models with unit roots (e.g. FRBUS). Changes: - Add has_unit_roots::Bool field to caches struct with proper reset - Detect unit roots from QZ eigenvalues in Schur QME path - Detect unit roots from eigvals(sol) in doubling QME path (solution.jl) - Add fast-path in Lyapunov dispatcher: skip directly to Schur deflation - Thread has_unit_roots flag through moments.jl, get_functions.jl, rrules.jl - Safety net: NaN detection in calculate_covariance as final fallback - Skip unnecessary Schur in variance decomposition for non-unit-root models Verified: RBC (no unit roots, all finite), FRBUS (89 finite, 339 NaN, flag set during first-order solve before covariance call). Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * Rename _detect_unit_roots, _ordered_schur! to drop leading underscore Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> --------- Co-authored-by: Thore Kockerols <Thore.Kockerols@ecb.europa.eu> Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
…nd update related function signatures Co-authored-by: Copilot <copilot@github.com>
… job parameter in LAPACK call; only one pair of schur vectors
…m selector based on problem size; update relevant function signatures for consistency. Co-authored-by: Copilot <copilot@github.com>
…on when workspaces are present
…ity for unit-root blocks Co-authored-by: Copilot <copilot@github.com>
…oved accuracy. less tringent -> 1-e7; should allow NAWM to converge using schur
…nchmark-only mode - Updated `extract_dynare_results.m` to include a benchmark-only mode for selective result extraction. - Modified `generate_julia_results.jl` to support benchmark-only models and improved argument parsing for thread configuration. - Enhanced `run_all_dynare.sh` to enforce single-thread execution in CI and added environment variable configurations for thread limits. - Introduced `run_all_dynare_windows.ps1` for Windows environments, implementing similar functionality as the shell script with support for thread counts and model skipping. - Added `run_thread_sweep_windows.ps1` to facilitate sweeping through different thread counts for performance benchmarking across models. Co-authored-by: Copilot <copilot@github.com>
…elling.jl into optim_LFI_alloc
…tation Co-authored-by: Copilot <copilot@github.com>
Add early return with NoTangent() in solve_sylvester_equation and solve_lyapunov_equation pullbacks when the adjoint solve fails. Previously the pullback continued computing gradients from the failed solve result, producing garbage ∂A, ∂B, ∂C values. Check Lyapunov solved flag in the Kalman filter rrule forward pass and return on_failure_loglikelihood when it fails, instead of silently proceeding with an invalid initial covariance. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Rebuild the neutral NSSS solver cache seed when clearing solution caches so long workflows cannot retain stale evicted warm-start entries. Add regression coverage for final-horizon IRF parameter derivatives after forced NSSS solver-cache eviction. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
The 'Gali :first_order (square obs)' test failed because the chosen observables [:log_y, :log_W_real, :log_N] are linearly dependent in the first-order solution. The model has the static identity W_real = Y^σ·N^φ (from W_real = C^σ·N^φ and C = Y), which means the 3×3 shock Jacobian is rank 2 and the LU factorization correctly fails. Replace log_W_real with i_ann (annualized nominal interest rate), which is determined by the Taylor rule and provides an independent channel for the monetary policy shock. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
…elling.jl into optim_LFI_alloc
- Add Mooncake NUTS sampling (100 samples) alongside existing ForwardDiff - Add Mooncake vs ForwardDiff gradient comparison for combined objective - Add Mooncake vs ForwardDiff gradient comparison for full log posterior - Import Mooncake, DifferentiationInterface, AutoMooncake Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Replace min(norm(A*X²), norm(C)) normalization with the standard relative residual from numerical linear algebra: norm(A*X² + B*X + C) / (norm(A*X²) + norm(B*X) + norm(C)) The previous min-based normalization (introduced in a765b3e) picked the smallest of only two of three terms, making it degenerate when matrix term scales differ by orders of magnitude — common in OBC models with many auxiliary lag equations where norm(C) << norm(A*X²). The sum-of-terms denominator is robust to arbitrary scale differences and accounts for all three equation terms. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Tests SS, first-order solution, and IRFs for a model with max() operator (occasionally binding constraint), which exercises the QME solver with large scale differences between matrix terms. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Changes in src/get_functions.jl:
get_estimated_shocks, get_estimated_variables, get_model_estimates,
get_estimated_variable_standard_deviations, get_loglikelihood) for
reproducible outputs
- Fix get_statistics code: replace broken RBC.var with get_variables(RBC)
and remove crashing parameters=get_parameters(RBC) kwarg
- Update get_conditional_forecast: shock names now include ₍ₓ₎ subscript
(delta_eps₍ₓ₎, eps_z₍ₓ₎), column display updated for 80-char width
- Update get_solution: fix -0.0 to 0.0 in k₍₋₁₎ row
- Update get_solution(RBC, parameter_values): floating point values updated
- Update get_conditional_variance_decomposition: near-zero (:A) values
(5.88e-32, 9.65e-31) now display as 0.0
- Update get_variance_decomposition: near-zero (:A) value (9.78e-31)
now displays as 0.0
- Update get_statistics covariance output: truncated with … for 80-char width
Changes in src/inspect.jl:
- Update get_steady_state_equations: negation now parenthesized,
:(-z{TFP} * ...) becomes :(-(z{TFP}) * ...)
- Update get_nonnegativity_auxiliary_variables: now returns 3 elements
(➕₁, ➕₂, ➕₃) instead of 2
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
…blocks - Add silent=true to all @parameters calls in jldoctest blocks to suppress non-deterministic timing output and model summary from being captured - Add report_missing_parameters=false to doctests with intentionally incomplete parameter definitions (get_missing_parameters, has_missing_parameters) - Fix get_conditional_forecast doctest: move alternative input format examples out of # output section into separate julia code block - Fix get_statistics doctest: split multiple # output sections into tested jldoctest + illustrative julia blocks - Fix get_non_stochastic_steady_state_residuals doctest: split multiple # output sections; add filter for machine-epsilon residual instability - All 39 doctests now pass with doctest(MacroModelling) Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
…elling.jl into optim_LFI_alloc
…elling.jl into optim_LFI_alloc
thorek1
commented
May 13, 2026
|
|
||
| ## When to use calibration equations | ||
|
|
||
| Calibration equations are useful when an empirical target is easier to observe than the structural parameter itself. Common cases include: |
Owner
Author
There was a problem hiding this comment.
are usefeul for example when...
or simply when you want to pin down a steady state value
Owner
Author
There was a problem hiding this comment.
are usefeul for example when...
or simply when you want to pin down a steady state value
thorek1
commented
May 13, 2026
| K[ss] = K_ss | beta | ||
| ``` | ||
|
|
||
| ### Multiple calibration equations in one block |
Owner
Author
There was a problem hiding this comment.
i dont see the added value of this
Owner
Author
There was a problem hiding this comment.
i dont see the added value of this
thorek1
commented
May 13, 2026
|
|
||
| ## Common pitfalls | ||
|
|
||
| **No solution exists.** If the target is inconsistent with the model structure the steady-state solver will fail. For example, requesting a capital-output ratio that implies a negative depreciation rate has no valid solution. Review the target value and the model equations when this happens. |
Owner
Author
There was a problem hiding this comment.
pick another example. this one is presented as trivial but is actually not.
Owner
Author
There was a problem hiding this comment.
pick another example. this one is presented as trivial but is actually not.
thorek1
commented
May 13, 2026
|
|
||
| **No solution exists.** If the target is inconsistent with the model structure the steady-state solver will fail. For example, requesting a capital-output ratio that implies a negative depreciation rate has no valid solution. Review the target value and the model equations when this happens. | ||
|
|
||
| **Solver convergence.** The nonlinear solver needs a reasonable starting region. Providing bounds and initial guesses for calibrated parameters can help: |
Owner
Author
There was a problem hiding this comment.
say that in the absolute majority of cases there is no need to provide a guess because the ss solver will most likely find the ss regardless but if the ss solver fails a guess can still be provided and will be helpful nonetheless. reformulate this in a more elegant way
Owner
Author
There was a problem hiding this comment.
say that in the absolute majority of cases there is no need to provide a guess because the ss solver will most likely find the ss regardless but if the ss solver fails a guess can still be provided and will be helpful nonetheless. reformulate this in a more elegant way. same for bounds
thorek1
commented
May 13, 2026
|
|
||
| Bounds must be written as standalone comparison statements (e.g. `0 < α < 1` or `α > 0` on separate lines), not appended to the calibration equation. The `guess` keyword provides an initial value for the solver. | ||
|
|
||
| **One equation per parameter.** Each calibration equation pins exactly one parameter. Adding a second calibration equation for the same parameter, or using one equation for two parameters, will cause an error. |
Owner
Author
There was a problem hiding this comment.
you cant use one equation for two parameters but you can use one equation twice for two parameters and that wont work. reformulate accordingly
Owner
Author
There was a problem hiding this comment.
you cant use one equation for two parameters but you can use one equation twice for two parameters and that wont work. reformulate accordingly
thorek1
commented
May 13, 2026
| get_steady_state(model_name) | ||
| ``` | ||
|
|
||
| The output table shows all steady-state variable values and calibrated parameter values. Check that the targeted variables match the specified values. |
Owner
Author
There was a problem hiding this comment.
no need to check. the solver will say it failed, otherwise the calibration equations will be satisfied
Owner
Author
There was a problem hiding this comment.
no need to check. the solver will say it failed, otherwise the calibration equations will be satisfied
thorek1
commented
May 13, 2026
|
|
||
| ## When to use calibration equations | ||
|
|
||
| Calibration equations are useful when an empirical target is easier to observe than the structural parameter itself. Common cases include: |
Owner
Author
There was a problem hiding this comment.
are usefeul for example when...
or simply when you want to pin down a steady state value
| K[ss] = K_ss | beta | ||
| ``` | ||
|
|
||
| ### Multiple calibration equations in one block |
Owner
Author
There was a problem hiding this comment.
i dont see the added value of this
|
|
||
| ## Common pitfalls | ||
|
|
||
| **No solution exists.** If the target is inconsistent with the model structure the steady-state solver will fail. For example, requesting a capital-output ratio that implies a negative depreciation rate has no valid solution. Review the target value and the model equations when this happens. |
Owner
Author
There was a problem hiding this comment.
pick another example. this one is presented as trivial but is actually not.
|
|
||
| **No solution exists.** If the target is inconsistent with the model structure the steady-state solver will fail. For example, requesting a capital-output ratio that implies a negative depreciation rate has no valid solution. Review the target value and the model equations when this happens. | ||
|
|
||
| **Solver convergence.** The nonlinear solver needs a reasonable starting region. Providing bounds and initial guesses for calibrated parameters can help: |
Owner
Author
There was a problem hiding this comment.
say that in the absolute majority of cases there is no need to provide a guess because the ss solver will most likely find the ss regardless but if the ss solver fails a guess can still be provided and will be helpful nonetheless. reformulate this in a more elegant way. same for bounds
|
|
||
| Bounds must be written as standalone comparison statements (e.g. `0 < α < 1` or `α > 0` on separate lines), not appended to the calibration equation. The `guess` keyword provides an initial value for the solver. | ||
|
|
||
| **One equation per parameter.** Each calibration equation pins exactly one parameter. Adding a second calibration equation for the same parameter, or using one equation for two parameters, will cause an error. |
Owner
Author
There was a problem hiding this comment.
you cant use one equation for two parameters but you can use one equation twice for two parameters and that wont work. reformulate accordingly
| get_steady_state(model_name) | ||
| ``` | ||
|
|
||
| The output table shows all steady-state variable values and calibrated parameter values. Check that the targeted variables match the specified values. |
Owner
Author
There was a problem hiding this comment.
no need to check. the solver will say it failed, otherwise the calibration equations will be satisfied
Codecov Report❌ Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## main #296 +/- ##
==========================================
- Coverage 87.05% 86.92% -0.13%
==========================================
Files 23 35 +12
Lines 14260 22603 +8343
==========================================
+ Hits 12414 19648 +7234
- Misses 1846 2955 +1109 ☔ View full report in Codecov by Harness. 🚀 New features to boost your workflow:
|
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.