margEVT 0.5.0

Breaking changes

• The point-process likelihood now uses the [.]_+ convention of the integrated intensity: when xi_t < 0 and the upper endpoint mu_t - sigma_t / xi_t lies below the threshold, period t contributes zero to Lambda. Previously such parameter vectors were rejected (1e9) while the gradient already treated them as zero, so fits with a negative shape could stop at an artificial boundary without being stationary. Values remain 1e9 only outside the model support (an exceedance beyond the upper endpoint, or a threshold below the lower endpoint when xi_t > 0).
• New fit_nhpp(shape_lower = -0.5): the shape parameter is restricted to xi_t >= shape_lower (Smith, 1985, regular region; below -1 the likelihood is unbounded). For a shape design with an intercept and at most one covariate the bound is handled exactly, with KKT conditions and multipliers; otherwise it acts as a barrier. Use shape_lower = -Inf for the previous unbounded behaviour. Fitted objects store shape_lower and shape_at_bound, and print() reports an active bound.
• pp_nllh() and pp_grad() gain shape_lower (default -Inf).
• Bootstrap and backtest refits inherit shape_lower from the fit.
• fit_nhpp() gains scaling, with default "auto" (see Performance). Estimates change within the optimizer tolerance, and selected models were unchanged in all tests; scaling = "none" reproduces earlier versions.

Estimation

• New active-set solver for the shape bound: bounded shape coordinates at the bound are fixed there and the remaining coordinates are optimized by BFGS, with the active set updated from the multiplier signs. L-BFGS-B, used before, stalled near the smoothed L1 kink: on a daily temperature refit at the bound it ran ~250 s without passing the KKT check, against ~20 s and a certified (lower) optimum now. Fits that start at the bound (along a regularization path, or in bootstrap and backtest refits once the bound is active) go straight to this solver; L-BFGS-B remains a fallback.
• KKT check under the shape bound: the subgradient of a penalized shape covariate at zero is now chosen exactly (the violation is piecewise linear in it) instead of on a grid. With adaptive weights (lambda * w of order 1e5) the grid missed the admissible interval and optimal fits were reported as non-converged, leaving gaps in the regularization path (e.g. 27 of 56 path points for a daily temperature fit).
• When the KKT check fails, BFGS is first restarted from its solution (fresh Hessian approximation) before falling back to L-BFGS-B; this polishes fits that stop just short of the tolerance and removes most failed path points.

Performance

• The likelihood is evaluated in a single pass (z_u, log z_u and tau once; z_u^(-1/xi - 1) as tau / z_u), about 4 times faster per call.
• The objective passed to optim() computes the gradient only when it is requested: line-search trials, many of which fall outside the support, now cost one value evaluation. Previously every call also computed the gradient, and support violations went through the slow path.
• Together: an unpenalized fit on 1000 observations went from ~60 ms to ~19 ms, and an adaptive-lasso fit with the BIC path from ~5.0 s to ~1.4 s, with the same estimates (differences of order 1e-8). The previous implementation is kept as .nhpp_eval_ref() for support violations with gradient and for Gumbel limits (|xi| < 1e-6).
• fit_nhpp() sets options(matprod = "blas") while it runs (restored on exit). Its covariates are checked to be finite, so the NaN scan that the default runs over the whole design matrix before each product is redundant; results are bitwise identical. On a daily-size adaptive fit (16436 days, 52 parameters) this cut the time from 182 s to 146 s.
• New argument scaling in fit_nhpp(). With "auto" (default) the optimizer receives the parameter scales 1 / sqrt(|H_kk|), from the diagonal curvature of the unpenalized likelihood at the starting values (computed once per call, reused along the regularization path and stored in the fit, so bootstrap and backtest refits use the same setting). Unscaled BFGS evaluated the objective about 9 times per gradient, cutting back steps of the wrong size. On the daily-size fit above: 147 s -> 77 s, same lambda and active set, lower objective (negative log-likelihood 1093.94 vs 1094.09) and a KKT violation 28 times smaller (6.8e-6 vs 1.9e-4), with all 54 path points certified (53 before). Estimates change within the optimizer tolerance; along flat directions of the likelihood the change can reach a few percent of a coefficient (2.6% for one location slope in that example), always towards a lower objective. scaling = "none" reproduces earlier versions bitwise.
• Benchmark of the JSS article (unpenalized fit, n = 1000): 62 ms -> 17 ms (extRemes 47 ms, ismev 176 ms on the same machine).

Bootstrap

• bootstrap_rl(), bootstrap_coef() and bootstrap_nhpp() gain occurrence: "fixed" (default, as before) keeps the observed number and timing of exceedances; "simulated" also draws them from the fitted point process (probability 1 - exp(-tau_t / n_y) per period), so the uncertainty about the exceedance rate enters the intervals.
• New bootstrap_nhpp(): return-level and coefficient intervals from one set of refits. New refit_nhpp() and annual_prob() (exported helpers).
• Arguments are evaluated before replicates are sent to worker processes (lazily evaluated arguments failed on PSOCK workers).

Backtesting

• backtest() gains approaches: "C" (default; resampled training years, as before) and/or "B", a VAR generator fitted to the training years only (generator_args are passed to fit_var_generator()). Both approaches are evaluated on the same window refit; results carry F_ann_<a> and exc_<a>_T<TR> columns, binom_tests and ks_test an approach column. The legacy columns F_ann and exc_T<TR> repeat the first approach.
• backtest() gains decluster, a function applied to the training-year series in each window, so declustering no longer uses validation data.

Minor

• build_cov_annual() no longer warns about a covariate missing from cov_vals when all its operational coefficients are zero (filling it with 0 is then exact). This happens whenever the VAR generator, which models only the active covariates, feeds trajectories to a fit with inactive candidates.

margEVT 0.4.0

Breaking changes

• fit_nhpp(lambda_scaling = ) now defaults to "common" (one scalar penalty for all parameter blocks), the convention used in the dissertation and articles. Use lambda_scaling = "gradient" for the previous score-calibrated block penalties.
• With lambda = "bic", the returned model is now the regularization-path solution at the selected penalty, so the reported fit is exactly the one scored by the BIC.
• Convergence is now certified by the Karush-Kuhn-Tucker conditions of the penalized problem (grad_tol, scaled by the number of exceedances) rather than by the optimizer’s return code. Fits that previously reported convergence may now warn, and vice versa.
• summary() returns an object of class summary.nhpp_fit (printed as before) instead of printing directly.
• The vars dependency was removed. VAR(p) generators are fitted by least squares with Schwarz order selection on a common sample, which reproduces vars::VARselect() and vars::VAR(); a single covariate is a VAR(p) with K = 1 rather than a separate AR model.
• Random-number streams changed: bootstrap replicates use independent L’Ecuyer-CMRG streams and trajectories are simulated jointly, so results for a given seed differ from 0.3.0.

Estimation

• Adaptive LASSO: penalty_weights = "adaptive" computes weights 1 / |theta|^adaptive_gamma from an unpenalized fit (ridge fallback); numeric coordinate weights are also accepted.
• New fit_nhpp() arguments lambda_grid, init, and grad_tol. The optimizer falls back from BFGS to L-BFGS-B and a restarted BFGS when the KKT check fails.
• Fitted objects store the penalty weights, the KKT violation, the evaluated regularization path (fit$path), and the observed series.
• The likelihood, gradient, and penalty are evaluated once per parameter vector and shared between objective and gradient calls.

Methods and diagnostics

• New methods logLik(), nobs(), predict(), residuals(), and coef() for nhpp_fit; AIC() and BIC() work through logLik().
• residuals() returns time-rescaling interarrival residuals and magnitude residuals, both standard exponential under the model.
• plot.nhpp_fit() gains type = "path" (BIC curve and coefficient paths) and type = "qq" (exponential Q-Q plots with exact pointwise bands).
• n_exceedances() uses the stored series when y is omitted.

Marginalization and covariates

• New seasonal_harmonics(); harmonics of any order (cos<k>, sin<k>, sen<k>) are recognized and rebuilt automatically.
• marginalize() returns an nhpp_rl object with Monte Carlo standard errors of the return-level roots (se_mc), accepts fixed covariate values (e.g. list(trend = 1)), fills inactive design columns silently, and has a plot() method.
• fit_var_generator() gains n_harmonics, excludes deterministic variables from the VAR, and supports derived covariates (lags, moving averages) through derive and derive_lead. simulate_covariates() simulates all trajectories at once and gains seed.

Bootstrap and walk-forward assessment

• Bootstrap refits start from the original estimate and keep the original penalty weights, gradient tolerance, and scaling. Results do not depend on the new cores argument; replicate matrices are returned as the attribute "replicates".
• backtest() gains trend_col (trend renormalized within each training window), fixed, reselect_lambda, and cores. Adaptive weights are recomputed inside each training window, and covariates inactive in a window refit are filled without warnings.

Documentation

• New vignette, examples for every exported function, and an updated citation file.

margEVT 0.3.0

Estimation consistency and reproducibility

• Added lambda_scaling to fit_nhpp(). Automatic BIC selection can retain the score-calibrated block penalties introduced in 0.2.0 or use a common scalar penalty to reproduce the estimation convention used in the dissertation analyses.
• Corrected automatic BIC selection so that both the unpenalized likelihood and the effective parameter count are evaluated at the thresholded operational estimator. Smooth optimizer solutions, rather than thresholded vectors, are retained as warm starts along the regularization path.
• Fitted objects now store both the smooth optimizer solution (par) and the operational activity-thresholded solution (par_oper). Downstream parameter prediction uses the operational estimator by default, while coef(fit, operational = FALSE) preserves access to the smooth solution.
• Coefficient bootstrap replicates now use the operational estimator, allowing percentile distributions to retain an atom at zero.
• Univariate covariate generators now fit a genuine stable AR model selected by BIC instead of padding the series with an artificial noise coordinate.
• Covariate simulation now supports exact Gaussian stationary-state initialization, while retaining zero-state burn-in as an explicit sensitivity option.

Validation and defensive checks

• Return-level integration now rejects non-finite fitted parameters and non-positive scales instead of silently omitting invalid contributions.
• Parametric and empirical marginalization verify all direct and interaction input covariates before constructing annual trajectories.
• Prediction matrices are rebuilt directly from the stored design columns, avoiding formula reinterpretation of non-syntactic or interaction-like column names.
• Package metadata and the dissertation citation were synchronized for the 0.3.0 development line.

margEVT 0.2.0

Consistency and temporal resolution

• Added active_tol to fit_nhpp() and stored it in fitted objects. The same tolerance is now used by automatic BIC selection, print(), summary(), bic_nhpp(), active_covariates(), bootstrap refits, and downstream covariate generators.
• build_cov_annual(), fit_var_generator(), simulate_covariates(), marginalize(), and backtest() now inherit the temporal resolution from the fitted model unless the user supplies an explicit override.
• Added frequency-aware behavior for weekly and other non-daily analyses.
• fit_var_generator() now verifies the VAR root condition, stores the root moduli and spectral radius, and refuses to simulate from an unstable fit.
• simulate_covariates() now generates independent annual Monte Carlo paths, each with its own burn-in, instead of splitting one dependent long path into nominal annual replicates.
• Corrected GEV endpoint handling in marginalized probabilities: levels below the lower endpoint when the shape is positive now have zero non-exceedance probability, while levels above a finite upper endpoint retain probability one conditionally on that trajectory.

Inference and documentation

• Backtesting now reports not_rejected alongside the legacy calibrated field and uses language appropriate for nominal binomial comparisons and exploratory tail-conditional PIT diagnostics.
• Validation years with insufficient observed coverage are now excluded from binomial denominators instead of being silently counted as non-exceedances.
• Approach C now uses a locally preserved random seed, preventing its annual block sampling from resetting the outer bootstrap stream and repeating subsequent parametric bootstrap draws.
• bootstrap_rl() now reports the original fitted-model return level in RL_est; previous versions reported the mean of successful bootstrap roots.
• Clarified that the bootstrap is conditional on the observed occurrence pattern and selected model specification.
• Corrected the documentation of empirical annual-block marginalization and the event-based BIC convention.
• Added strict validation for block-specific penalty controls, return periods, Monte Carlo samples, and covariate trajectory lengths to prevent silent recycling or malformed fits.
• Return-level root finding no longer expands below the fitted threshold. It expands only the upper bracket and returns NA when the requested quantile lies outside the point-process tail domain.

margEVT 0.1.1

• Added plot.nhpp_fit() for cumulative-intensity and fitted-parameter plots.
• Added package URL and bug-report metadata.

mirror server hosted at Truenetwork, Russian Federation.