NEWS.md

In qpmR: Quarterly Projection Models for Monetary Policy Analysis

qpmR 1.1.0

Speed

• A compiled (C++/RcppArmadillo) Kalman filter, used by default. It is a line-for-line match of the reference implementation kept in R, down to the Cholesky factorisation and Joseph-form update, and the two are tested to agree to machine precision on real data, with missing observations, with measurement error, and on the collinear case that raises qpm_singular_F. Set options(qpmR.use_cpp = FALSE) to force the R path; qpm_use_cpp() reports which is in use.
• The stationary covariance is now obtained by squaring rather than by vectorising to an N^2 x N^2 system — O(N^3) per iteration instead of O(N^6). This turned out to matter more than the filter: it is 39x faster on the Czech model and also speeds up model_properties(), qpm_rule_eval() and qpm_identify(), which all solve the same equation.
• The parameter-independent structure of the first-order system — which auxiliary states are needed, the expanded state vector, and the exact cell of A, B, C or D that each coefficient lands in — is now computed once per model and cached on it. Estimation re-solves the same equations thousands of times with different numbers, so only the values are recomputed: build_first_order() went from 27.5 ms to 1.7 ms (16x) and qpm_solve() from 37 ms to 8.3 ms. The cache is rebuilt automatically if it is missing (a model serialised by an older version) or no longer matches its equations. Linearity is checked once at construction rather than on every solve, since it is a property of the equations.
• Coefficient extraction reuses one evaluation environment per equation instead of rebuilding it per symbol.
• Together these take a posterior draw on the Czech model (22 states, 110 quarters) from 181 ms to 27 ms — 6.7x end to end, or a 6000-draw estimate from about 18 minutes to under 3. The compiled filter is now the largest single cost, which is where the time should be: the QZ decomposition is 3.5 ms and model assembly 1.7 ms.
• The eigenvalue table behind eigen_table() is assembled directly from the QZ output rather than through data.frame(), whose constructor, reorder and row-name reset were about a fifth of a posterior draw on small models even though the estimation objective never reads the table. The table itself is unchanged.

New features

• qpm_risk() / risk_log(): express a balance of risks. Bands become two-piece normal, so the mode stays on the model's projection while the mean shifts by the stated skew and total variance is held fixed — a skew redistributes risk rather than adding it. Fan charts centred on the mode, as published fan charts are. Unlike add_judgment(), which moves the projection and back-solves the supporting shocks, this changes only the shape of the distribution around an unchanged path.
• qpm_rule_eval(): score alternative policy rules over a grid by the unconditional loss, computed exactly from the stationary covariance rather than by simulation, and trace the inflation-output variability frontier. Rules that fail Blanchard-Kahn are reported as indeterminate or explosive rather than silently dropped.
• qpm_counterfactual(): replay history with shocks switched off or scaled — "what if the central bank had simply followed its rule?". Replaying the unmodified shocks reproduces the smoothed history to machine precision, which the function checks.
• qpm_compare_models(): compare model behaviour (impulse responses and implied moments), complementing qpm_diff(), which compares structure.
• qpm_disaggregate(): temporal disaggregation of annual data to quarterly by Denton-Cholette or Chow-Lin, both satisfying the aggregation constraint exactly — the first step for the many economies that publish national accounts only annually.
• fevd(): forecast error variance decomposition — how much of each variable's forecast uncertainty each structural shock accounts for, at every horizon. Shares sum to one by construction (verified to 1e-10 in the tests), exogenous processes come back as entirely own-driven, and the decomposition is well defined for unit-root models even though the variances themselves are not.
• model_properties(): the standard calibration check — model-implied standard deviations and autocorrelations from the stationary covariance, the shock dominating each variable's unconditional variance, and the same statistics computed from data alongside, with a warning when model and data volatility differ by more than a factor of two.
• Standard R generics on qpmR objects, so nothing has to be reimplemented: logLik() and nobs() for filtrations and estimates (which makes AIC() and BIC() work), residuals() (one-step innovations, standardised innovations, or smoothed structural shocks) and fitted() for filtrations, vcov() and confint() for estimates, and summary() methods returning data frames for both.
• write_dynare(): export any model as a Dynare .mod file. The original equations are exported rather than qpmR's internal first-order system, so Dynare builds its own auxiliary variables for long leads and lags and the two implementations agree only if both are right. The test suite now pins qpmR's impulse responses to golden files generated by Dynare 6.0: across 4080 points (12 shocks x 17 variables x 20 quarters) the largest discrepancy is 1.4e-14, and steady states agree to 8e-14. The food-block model agrees to 1.7e-14, which independently validates add_block(). Regenerate the golden files with data-raw/dynare_golden.R.
• A pkgdown website, with the reference index organised by workflow stage rather than alphabetically.
• Test coverage is measured on every push and reported to Codecov.

Smaller changes

• Printing a solved model whose stable roots are all unit roots (a pure random walk) no longer reports -Inf as the largest stable root.
• Conditional fan bands (qpm_condition(), add_judgment(), scenarios) are computed from the conditioned rows and the diagonal of the stacked-path covariance rather than from the full (H N) x (H N) matrix, which was the package's one large matrix product and imposed a size cap above which bands fell back to the unconditional ones. The cap is gone; band values agree with the previous formula to 1e-9.
• write_dynare() no longer writes a file by default: file = NULL (the new default) returns the Dynare source as a character vector, and a file is written only when a path is given. qpm_report() and save_round() likewise require an output path and a store directory rather than defaulting to the working directory.
• qpm_estimate() reports progress with message() rather than cat(), so suppressMessages() silences it.
• A seed given to simulate() or qpm_estimate() no longer disturbs the caller's random number stream: the previous RNG state is restored on exit, as stats::simulate() does for linear models.
• The Description cites the Berg, Karam and Laxton (2006) how-to guide by DOI, and the slower \donttest{} examples were resized to run in a few seconds each.

qpmR 1.0.0

First stable release. The full forecasting-and-policy-analysis workflow now runs end to end — data, filtering, gaps, model, baseline, judgment, policy, scenarios, rounds, revisions, verification, report — and is exercised on a real quarterly dataset for Czechia shipped with the package.

This release adds the reporting and audit layer and country-adaptation blocks on top of 0.4.

Reporting and audit

• qpm_report(): turns a round into the document a policy meeting is run from — executive summary with the numbers filled in, forecast table and fan charts, filtered gaps, shock decomposition, the judgment ledger with its implied shocks, an optional revision decomposition against the previous round, and a reproducibility appendix. The .Rmd source is always written (institutions replace the template's text, not its plumbing) and rendered to HTML/PDF/Word when pandoc or Quarto is available; where neither is — air-gapped forecasting machines, bare CI runners — it says so and returns the source rather than failing.
• chart_pack(): the standard round chart set (forecast fans, filtered latent states, shock decomposition, monetary transmission) as a multi-page PDF or numbered PNGs.
• verify_round(): re-runs an archived round from its own contents and checks that the published numbers come back, reporting the largest deviation, the worst variables, and any qpmR version drift. When the round is loaded from a store it also checks the human-readable CSV sidecars against the object, so a hand-edited audit trail is detected.
• Stacked-bar decomposition and revision charts leave headroom for their legends.

Country adaptation

• qpm_block() / add_block(): reusable bundles of variables, shocks, parameters and equations that adapt a template to a country without forking it. Equations whose left-hand side names an existing variable replace that variable's equation; equations for newly declared variables are appended. Blocks compose, and each one is recorded in the model's meta$blocks.
• block_food_cpi(): headline CPI split into food and core. The Phillips curve moves to core, food gets its own persistence, stronger exchange-rate pass-through, and error correction on the relative food price, and headline becomes the weighted identity. Food is 30-50 percent of the basket across most of sub-Saharan Africa and South Asia, where a single-inflation model is unusable; a food supply shock in this block raises headline while leaving core essentially untouched, which is the relative-price story policy should look through.
• block_fx_intervention(): a leaning-against-the-wind intervention rule entering the UIP block, so one model spans a continuum of exchange-rate regimes — intensity = 0 reproduces the free float exactly, moderate values a managed float, large values approach a peg (the peak exchange-rate response to a risk-premium shock falls from 0.97 to 0.53 to 0.08 across those settings).
• qpm_template() gains the "bkl_food" and "managed_fx" shortcuts.
• qpm_diff(): structural comparison of two models — variables, shocks, parameters and equations added, removed or changed, plus recalibrations — so a country team's customization is reviewable as a diff rather than a fork.

qpmR 0.4.0

The estimation layer is complete.

Identification, marginal likelihood, vignette

• qpm_identify(): Iskrev-style local identification diagnostics before any sampling — numerical Jacobians of the solved model (solution level) and of the observables' population moments (moment level, stationary models) with respect to the chosen parameters. Reports parameters with no effect, rank-deficient combinations, and near-collinear pairs that are only jointly identified. Unit-root models get the solution-level check with an explanatory note.
• marginal_likelihood(): log marginal likelihood by the modified harmonic mean (Geweke 1999) across truncation probabilities with a stability spread, plus a Laplace approximation at the mode in transformed space as a cross-check. Differences across models on the same data are log Bayes factors. truncate() priors are now renormalized numerically at construction so they contribute proper densities.
• New vignette qpmR-estimation: priors, the AR(1) estimation laboratory, identification, Bayes factors, and the full Czech estimation with its results discussed (including the honestly weakly-identified policy-response coefficient).

Priors, sampler, posterior forecasts

• priors(): the prior mini-language. normal(), beta(), gamma(), invgamma(), uniform(), and truncate() exist only inside priors() (evaluated in a controlled environment), so base R's beta() and gamma() functions are never masked. Beta/gamma/ inverse-gamma use the mean/sd parametrization economists write down.
• qpm_estimate(): Bayesian estimation of any subset of structural parameters and shock standard deviations over the Kalman-filter likelihood. Posterior mode in transformed (unconstrained) space, BFGS Hessian as the proposal seed, adaptive random-walk Metropolis (Haario-style covariance adaptation during burn-in, acceptance targeted at 0.25), multiple sequential chains, split R-hat and Geyer effective sample sizes. Draws violating Blanchard-Kahn get zero weight (the usual determinacy truncation). method = "mle" reuses the machinery with flat priors on the declared supports.
• Printing reports mode, posterior mean, 90% interval, R-hat, ESS, and a "learned" column comparing posterior to prior spread -- a cheap identification diagnostic. plot() overlays prior and posterior densities. coef() extracts point estimates; apply_estimate() recalibrates the model at them.
• posterior_forecast(): fan charts integrating over the posterior -- each draw re-solves the model and re-filters the data, so the bands combine future-shock and parameter uncertainty.
• Internal: kalman_loglik(), a storage-free filter pass for estimation speed.

qpmR 0.3.0

The policy-analysis layer is complete.

Forecast rounds and revision decomposition

• qpm_round(): one replayable artifact per forecast -- model, calibration, data vintage, filtration, and the conditioned forecast together. [qpm_condition()], qpm_scenario(), and add_judgment() apply to rounds directly.
• save_round() / load_round() / list_rounds(): a plain-directory round store; each round is a self-contained round.rds plus human-readable CSV sidecars (forecast, data, calibration, judgment) for auditing without R.
• compare_rounds(): the revision decomposition. The forecast revision between two rounds is split into parameters, data revisions, new data (outturns), conditions, and judgment by re-running the full pipeline swapping one ingredient at a time. Contributions telescope (they sum to the total exactly); the endpoints are verified against the archived rounds, so a version drift is reported rather than silently absorbed. Judgment overtaken by data (a conditioned quarter that has become an outturn) is dropped and reported. Waterfall printing and stacked revision charts.
• next_quarters() exported for quarter-label arithmetic; formulas in models are stored without environments, keeping serialized rounds small.

Conditional forecasts, scenarios, judgment

• qpm_condition(): hard conditional forecasts. Impose paths on any variables at any horizons; qpmR backs out the minimum-norm structural shocks (in standard-deviation units, optionally restricted to instruments) that deliver them. The anticipated switch is explicit: TRUE means the conditioned path is announced at the start of the forecast and expectations react ahead of it, FALSE means period-by-period surprises. Anticipated propagation uses the exact news recursion F_j = N^j Q, N = -(AP+B)^{-1} A, verified in the tests against a brute-force perfect-foresight solve. Fan bands are recomputed as the Gaussian conditional distribution given the conditions (zero width at conditioned points). Announced rate holds reproduce the Laseen-Svensson (2011) anticipated-path reversal, as they should.
• qpm_scenario(): shock-based alternative scenarios (announced or surprise), additive on any forecast.
• add_judgment() / judgment_log(): the judgment ledger. State the adjustment in percentage points; qpmR back-solves the supporting shocks, keeps the forecast model-consistent, records author, timestamp and rationale, and flags judgment requiring shocks above two standard deviations. Entries are stored as absolute targets and the full condition/judgment set is re-solved jointly, so the ledger is replayable.
• Forecasts carry quarter labels (2026-Q3 style) inherited from the filtration; conditions and judgment can be addressed by label or by horizon (h3). Conditioned and judgment points are marked on fan charts.

qpmR 0.2.0

The filtration layer is complete.

Kalman filter, smoother, decompositions

• qpm_filter(): Kalman filter + RTS smoother over the solved model. Jointly infers every latent state (output gap, neutral rate, equilibrium exchange rate, trends) and the historical structural shocks from any subset of observed variables, with missing data and ragged edges handled naturally. Innovation diagnostics (Ljung-Box, outlier flags) are computed and printed. The likelihood is tested against the exact closed-form Gaussian likelihood.
• qpm_decompose(): exact historical shock decompositions of the smoothed history (additivity verified internally), with stacked-bar plots.
• state_space(): exports the exact T, R, Z, H, Qc, P1 matrices used internally, so other estimators can build on qpmR.
• qpm_forecast() accepts a qpm_filtration as from, forecasting from the smoothed end-of-sample state with the smoothed history kept for fan charts.

Diffuse initialization and unit-root trends

• Models with unit roots (random-walk trends) now solve: roots within unit_tol of the unit circle count as stable (the usual qz-criterium convention) and are reported separately. Steady states with free trend levels use a minimum-norm least-squares normalization; a drifted random walk (no fixed point) is a typed qpm_no_steady_state error explaining the balanced-growth limitation.
• qpm_filter()/state_space() switch automatically to an approximate diffuse initialization (damped-Lyapunov large-variance prior, kappa = 1e6) when the model has unit roots. Exact Durbin-Koopman diffuse recursions remain on the roadmap.
• qpm_template("bkl", trends = "rw"): equilibrium real exchange rate and potential growth as driftless random walks; the neutral rate stays anchored by real interest parity (a free random walk there would make steady-state gaps indeterminate).
• The template gains a GDP-growth observation block (dy_obs = dy_bar + 4 * (y_gap - y_gap[-1])), so the model filters on actual national-accounts data without modelling the level of potential output.

Example country dataset

• czechia: quarterly Czech data 1996Q1 onward in model units (CPI inflation QoQ and YoY, 3M PRIBOR, real CZK/EUR, GDP growth, EURIBOR, euro-area HICP), compiled reproducibly by data-raw/czechia.R from FRED/OECD/Eurostat/ECB public endpoints. Filtering it with the rw template reproduces the known history: the pre-GFC boom, the 2009 and 2013 recessions, the COVID crater, the koruna's trend real appreciation, and the post-GFC fall in potential growth (pinned in the test suite).

Smaller improvements

• Filtration and decomposition charts label the time axis with period labels; plot() on decompositions gains a periods window argument.
• Extended qualitative palette (no more colour recycling with many shocks).

qpmR 0.1.0

• Initial release: model DSL (qpm_model(), x[-1] / E(x[+1]), automatic auxiliary states), Klein/QZ solver with Blanchard-Kahn diagnostics, steady states, irf(), simulate(), qpm_forecast() with fan bands, qpm_lint(), and the canonical Berg-Karam-Laxton small-open-economy template qpm_template("bkl").

qpmR documentation built on Sept. 29, 2026, 5:10 p.m.