thorek1/MacroModelling.jl: v0.2.0

MacroModelling v0.2.0

Diff since v0.1.46

changelog: Major internal refactor and many new features

User-facing

Estimation, filtering, and smoothing

• Filter-free joint loglikelihood estimation.
• Missing-value support in estimation, filtering, and smoothing.
• User-provided initial_state support in estimation.
• Fixed a correctness issue in the second-order inversion filter.
• Added warmup iterations to inversion-filter loglikelihood rrule.

Statistics, moments, and decompositions

• Added higher-order (pruned 2nd/3rd) variance decomposition.
• Added correlation to get_statistics, including derivatives.
• Added marginal_contribution (Shapley) shock decomposition for the inversion filter and marginal_contribution variance decomposition at pruned 2nd/3rd order, using a polynomial-coefficient algorithm (∑_{j≤2k} C(nᵉ,j) Lyapunov solves) in place of the 2^nᵉ exhaustive coalition enumeration, making higher-order Shapley aggregation feasible for models with many shocks.

Solutions and IRFs

• get_irf with parameters also works at higher order.

API, model writing, and Dynare interop

• Custom steady-state function support (including in-place functions).
• get_equations and get_calibration_equations accept a filter keyword argument.
• write_mod_file (Dynare) allows modifying order, pruning, and IRF length.
• Equation modification supported (Troll-style).
• Counters for steady-state and perturbation solves.

Autodiff and ecosystem compatibility

• All functions intended for use with derivatives now work with forward- and reverse-mode AD across all options (higher order, filters, etc.).
• Compatibility with Mooncake (and switched the docs to Mooncake).
• Compatibility with Turing <=0.45.

New models

• Added the FRB/US model.

Internal

Performance and allocations

• Overall greatly reduced allocations and improved performance.
• Preallocated BLAS/LAPACK calls (QZ, LU, QR) throughout.
• Preallocated workspaces for all solvers, with better reuse; uses LinearSolve and FastLapackInterface for matrix solves.
• Many indices precomputed; constants moved to a separate struct; workspaces and caches reused across repeated solves.
• AD paths of the doubling algorithms sped up by caching matrix powers.
• Custom sparse-matrix kernel for higher-order solves.

Solutions and Solvers

• Higher-order solution algorithms moved to compressed space (exploiting symmetry).
• QME solver switches to doubling for large systems by default.
• Analytical OBC Jacobian replaces ForwardDiff.
• Bartels–Stewart Sylvester and Lyapunov solvers moved from MatrixEquations.jl to an extension.
• Large Sylvester systems solved with Krylov methods now benefit from an ILU preconditioner.
• Lyapunov solver: accepts an initial guess, supports early termination, better handles unstable systems, the Krylov solver now operates on the upper-triangular system (implicitly enforcing symmetry of the solution), and adds dqgmres support.
• Lyapunov solver supports a has_unit_roots parameter for unit-root covariance handling.
• Much more detailed tolerance settings across solvers, with more robust defaults.

Architecture and refactors

• ForwardDiff is now an extension; internal derivatives use rrules based on analytical derivatives.
• NSSS solver refactored: struct dissolved into constants, functions, caches, and workspaces; solve_nsss_wrapper introduced as the API layer.
• DispatchDoctor type-stability coverage expanded across numerical source files.
• Removed direct dependencies on RecursiveFactorization and DifferentiationInterface.
• Equations are returned as expressions instead of strings.

Tests

• System-prior estimation is now tested.
• Replaced MCMCChains with FlexiChains in estimation tests.

Merged pull requests:

• Filter autodocs to show only exported functions (#255) (@Copilot)
• Optim_LFI_alloc (#262) (@thorek1)
• new code space dev container (#274) (@thorek1)
• refactor loglikelihood rrule conditions to simplify state updates; fi… (#275) (@thorek1)
• Bump actions/upload-artifact from 6 to 7 (#276) (@dependabot[bot])
• Bump julia-actions/cache from 2 to 3 (#282) (@dependabot[bot])
• Bump codecov/codecov-action from 5 to 6 (#284) (@dependabot[bot])
• Copilot/unit root covariance (#287) (@thorek1)
• Higher order decompositions (#291) (@thorek1)
• Cache squared matrices in doubling algorithms for AD speedup (#294) (@thorek1)
• Filter with missing values (#295) (@thorek1)
• Add filter-free estimation across all perturbation orders (#297) (@thorek1)
• Warmup (#298) (@thorek1)