Corrects shelf_life(), which could report a forecast
horizon that rested on a single noisy period. Any reported shelf
life should be re-checked; series whose CI-width /
response-scale ratio degrades monotonically are unaffected.
shelf_life() reported the first threshold
exceedance as the horizon. When the ratio hovers near
threshold without trending, one period can exceed it by
chance, and that isolated excursion was promoted to a “shelf life”.
Observed on a real series whose ratio sat at ~0.92 across a 16-period
window with a single excursion to 1.038: the reported horizon was not
reproducible across sampler settings, varying between “no crossing”, a
projection far outside the window, and two different crossing periods,
from MCMC noise alone.
shelf_life() now requires the exceedance to
persist. The new min_run argument (default
2) sets how many consecutive uninformative periods
constitute a crossing; min_run = 1 restores the previous
behaviour. A crossing already in force at the first lead is still
detected, and a dip inside an otherwise degrading series no longer
resets the horizon.
The projection mode extrapolated slopes indistinguishable
from flat. min_slope_for_projection is a magnitude
gate only, so a slope of ~1e-3 with a one-sided p-value of 0.42 cleared
it and was extrapolated to a confident-looking crossing time. On the
same real series above this produced a projected horizon 45 years out
whose delta-method 95% interval spanned 328 years and included
the past. The projection now additionally requires the slope to
be significantly positive, controlled by the new
projection_alpha argument (default 0.05); when
it is not, the result is reported as a lower bound — the honest outcome
for a forecast whose precision does not degrade in-window. Set
projection_alpha = 1 to restore the previous behaviour.
The two fixes compose: a series that neither sustains a crossing nor
trends detectably now returns lower_bound instead of either
a spurious observed horizon or a spurious projection.
et_skill_score() had the same first-crossing
defect in its forecast limit. The null-relative forecast limit
was the first lead time with skill <= 0, with no
persistence requirement. Skill hovering near zero crosses by chance, so
one unlucky lead was promoted to “the horizon” — found on a real
comparison whose mean skill over the window was comfortably positive.
et_skill_score() now takes min_run (default
2, matching shelf_life()) and the returned
"forecast_limit" attribute carries n_below,
first_below and min_run, so isolated dips are
visible rather than load-bearing. min_run = 1 restores the
previous behaviour.
Group-level variance channel for hierarchical
models. The decomposition now handles
(1 | g)-style fits correctly, and distinguishes the two
budgets that a hierarchical model implies:
param_var, and there is no separate channel. Conditioning
on a known group reduces predictive variance relative to the
population level, so the naive Var(full) - Var(population)
is negative and would be meaningless as a “group variance” (measured on
a 6-group fixture: 0.022 vs 0.244).group_var column
appears, carrying the between-group variance. Verified to recover the
fitted tau^2 on a known-truth fixture.et_predict() detects new levels automatically and
reports which budget it is computing.
param_var + group_var + env_var + residual_var (+ temporal_var)
still equals total_var exactly.
et_shelf_life_pool() — pool shelf lives
across forecast origins. A horizon from a single forecast
origin is a property of that training window as much as of the system.
This pools horizons from repeated origins and returns a median with an
interval. Origins where the forecast was still informative when the data
ran out are right-censored, not missing: discarding
them biases the pooled horizon downward, because the censored origins
are exactly the ones with long horizons. Estimation is Kaplan–Meier via
survival (a Recommended package, in Suggests),
with a documented, deliberately interval-free fallback when it is
unavailable. All origins censored is reported as a result — the forecast
never became uninformative — not as an estimation failure.
et_priors_split() — remove the
prior/likelihood double use entirely.
shrinkage = "zero" already stopped reusing point estimates
as prior means, but the prior scale, and any variable
selection performed by the first-stage model, still came from
the rows that form the likelihood. This fits the first stage on a random
(optionally stratified) subset and returns the disjoint complement for
the Bayesian fit, so selection and likelihood share no observations and
no post-selection caveat applies. The cost — a smaller likelihood sample
— is documented rather than hidden. The user-supplied seed is restored
afterwards so it cannot leak into the caller’s RNG stream.
Exceedance diagnostics on every horizon. The
horizon attribute now carries n_exceedances,
frac_exceedance, first_exceedance and
min_run alongside value and type,
so an isolated excursion is visible rather than silent. Comparing
first_exceedance with value shows directly
whether a horizon rests on a persistent crossing. Trend-mode horizons
additionally carry slope and slope_p.
Corrects the variance decomposition for models carrying a residual
autocorrelation term. Any analysis using ar() /
ma() / arma() / cosy() /
unstr() / sar() / car() must be
re-run; iid models are unaffected (all three changes below are
no-ops without an autocorrelation term). Note that 1.2.1 ships the
defect.
env_var was silently destroyed for any model
with an ar() / ma() / arma()
term. Under brms’s default ar(cov = FALSE)
parameterisation the autocorrelation term enters the
mean (mu[n] += ar * err[n-1]), not only
the covariance. decompose_uncertainty() computed
env_var = Var(mu_perturbed) - Var(mu_draws) where
mu_perturbed came from an internal hand-rolled linear
predictor (no AR term) while mu_draws came from
posterior_linpred() (AR term included). The difference was
therefore V_env - V_AR, which turns negative as the AR
error grows and was then clamped to zero by pmax(., 0).
In a controlled test (AR(1), phi = 0.9, analytic
V_env = beta^2 * delta^2 = 0.152) the reported
env_var averaged 0.012 and was floored to exactly
zero at 13 of 15 lead times; the iid control recovered 0.173
correctly. Both posterior_linpred() calls now pass
incl_autocor = FALSE, which returns the pure regression
linear predictor — verified bit-identical to the internal predictor, so
the AR term cancels exactly in the subtraction. env_var now
averages 0.145 with no floored rows.
param_var no longer absorbs the
autocorrelation error, and temporal_var no longer collapses
to zero. The same contamination meant param_var
grew with lead time even on non-trending new data (0.154 -> 0.976
over 15 steps, an 8.9x overstatement of parameter uncertainty at the far
lead), which in turn made
temporal_var = pmax(0, pp_var - (param_var + residual_var))
evaluate to zero at every lead for a series with phi =
0.82. After the fix param_var is flat (0.154 ->
0.147) and temporal_var grows monotonically 0 -> 0.637,
correctly reporting zero at lead 1 where a single innovation adds
nothing beyond residual_var.
This changes reported numbers for every autocorrelation
model. Analyses using
ar()/ma()/arma()/cosy()/unstr()/sar()/car()
must be re-run.
The env_var zero-floor is now
instrumented. decompose_uncertainty() warns when
the floor fires on rows whose shortfall exceeds twice its own Monte
Carlo SE — too systematic to be jitter. The silent floor is what hid the
bug above through two review rounds. The test is statistical rather than
a raw count, so small prediction sets (where a floored row genuinely is
jitter) do not trip it.
et_fit() gains ar_prior,
defaulting to "weakly_informative".
brms places a flat, unbounded prior on
class = "ar" / "ma" (get_prior()
reports (flat) with no lb/ub), so
nothing keeps the residual process inside the stationary region. On the
short series typical of ecological panels the posterior routinely
straddles the unit root, and the k-step forecast variance
sigma^2 (1 - phi^(2k)) / (1 - phi^2) then diverges. The new
default normal(0, 0.5) places ~95% of the prior mass inside
(-1, 1) without hard-bounding, so a genuinely
non-stationary series can still say so.
ar_prior = "stationary" truncates to (-1, 1);
ar_prior = "flat" restores the old behaviour. A prior the
caller supplies for these classes is never overridden.
On a near-unit-root fixture (n = 23, true random-walk residual) this reduces the ratio of raw to 1%-winsorized predictive variance from 1072x to 1.7x — i.e. the heavy tail that motivated the winsorization below was largely an artefact of the unbounded prior, not a property of the data.
et_fit() warns when more than 5% of an
autocorrelation posterior sits at |value| >= 1,
reporting the exact fraction and posterior mean. For AR(1) this is
precisely the posterior probability of a non-stationary process. The
message directs the user to read long-lead intervals,
temporal_var and shelf_life() as “no usable
horizon” rather than as calibrated numbers.
et_predict() gains var_trim
(default 0), replacing the always-on, undocumented 1%
winsorization. Two problems are fixed:
param_var and the environmental pair
stayed raw. temporal_var was therefore a difference between
two different estimators, biased low by whatever the winsorization
removed (~3.6% for a Gaussian at trim = 0.01) — which
blinded it to genuine autocorrelation shares below that threshold.
var_trim is now applied uniformly to every
channel.var_trim = 0 default they
describe the same one.The tail diagnostic is still computed (against a fixed 1% reference,
regardless of var_trim, since it is a statement about the
model rather than the estimator) and now recommends
ar_prior = "stationary" instead of implying that a robust
summary makes a diverging forecast trustworthy.
This changes reported variance components wherever the old winsorization was biting — i.e. any near-unit-root autocorrelation fit.
Regression tests added in
tests/testthat/test-decompose-autocor.R, covering
env_var recovery under ar(), floor-rate, no
decay with lead time, positive temporal_var, flat
param_var, and an iid control. Reproduction scripts in
tests/manual/. The previous suite missed all of this
because its mock builds mu_perturbed as
lp + noise, i.e. it assumed the two arrays were already on
the same footing.
v_env_mcse on aarch64
(macOS arm64). This is the failure that caused the CRAN
archival on 2026-06-08. decompose_uncertainty() detected
the “no environmental noise” case by testing the variance difference
v_env_raw == 0 exactly, and it computed the two column
variances with different stats::var() na.rm
settings. Those settings select different C code paths in
cov.c, which round identically on x86_64 but not on
aarch64. A last-bit difference therefore flipped the reported Monte
Carlo SE of env_var from 0 to a large value
(0.20 in the CRAN log) when env_noise = 0. The two
variances are now computed the same way, and the zero test uses a
relative tolerance instead of exact equality, so the result no longer
depends on bit-level floating-point reproducibility.
env_var itself was already correct to within 1e-10; only
its Monte Carlo SE was affected. Regression tests added.This release addresses the correctness issues raised in peer review, starting with the uncertainty-budget consistency of the decomposition.
Total variance now contains environmental uncertainty
(T1). decompose_uncertainty() previously computed
total_var from posterior_predict() on the
unperturbed predictors, so env_var was an add-on
that sat outside the total — the reported components could sum
to well over 100% of total_var. total_var is
now defined as the sum of its components (law of total
variance): param_var + env_var + residual_var for iid
models, plus temporal_var for autocorrelation models. The
percentage shares now sum to 100% exactly, and env_var is a
genuine sub-share of a total that contains it.
et_predict() folds environmental uncertainty
into the credible interval by default when env_noise is
supplied. The include_env_in_ci default changed
from FALSE to NULL (auto): env is now included
whenever the caller passes non-zero env_noise, so the
reported forecast interval — and therefore shelf_life() and
et_calibrate() coverage — reflect predictor/driver
uncertainty and stay coherent with total_var. Pass
include_env_in_ci = FALSE to recover the previous
parameter+residual interval. The env-inclusive predictive is now
constructed family-generally (re-centring each posterior-predictive
residual on the perturbed mean), rather than the previous Gaussian-only
lp + sigma * N(0,1).
et_sensitivity_profile() /
et_plot_sensitivity(): env_share is now
env_var / total_var (previously
env_var / (env_var + total_var)), which double-counted env
once env entered the total.
extract_priors() gains a
shrinkage argument, defaulting to "zero"
(T5). This addresses the prior/likelihood double-use objection
at its source. The previous behaviour — using the fitted coefficients as
prior means and then refitting the Bayesian model on the same
data — reused the data twice and could make posteriors overconfident.
The new default centres every coefficient prior at 0 (a regularizing,
ridge-style prior) with a width scaled by the coefficient
magnitude multiplier * |coef| (floored at
min_sd), so a well-estimated large coefficient is not
shrunk toward 0 while its estimated location is no longer
reused. Set shrinkage = "estimate" to recover the previous
informative-mean behaviour (with a documented double-use caveat). For
ranger the prior mean is always 0, so
shrinkage has no effect there. The
et_prior_spec object and its print() method
now report the shrinkage mode.
et_fit() now documents its boundary plainly (a thin
brms wrapper; the decomposition of hierarchical fits is
provisional pending a group-variance term, and grouping=
fits independent per-group models rather than one multilevel model), and
emits a warning on high Rhat / divergent transitions pointing to
et_diagnose() before prediction.
extract_priors.ranger() is documented and flagged
(advisory) as : random-forest importance has no sign or GLM-coefficient
scale, so its importance→prior-SD mapping is heuristic — prefer
glmnet/lm/glm for a principled
prior.AR/MA forecast variance now accumulates with lead time
(T3). For autocorrelation models
(ar()/ma()/arma()/cosy()/unstr()/sar()/
car()), et_predict() previously called
posterior_predict() on the forecast rows alone, which does
not propagate the residual autocorrelation across the
horizon — the predictive variance (and hence the credible interval,
temporal_var, and shelf life) stayed flat with lead time.
et_predict() now continues the training series with the
forecast rows (response blanked to NA) so brms iterates the
fitted process forward: the forecast variance grows with lead time,
integrates over the posterior of the AR coefficient (draws with |φ| ≥ 1
contribute super-linear growth, correctly handling the non-stationary /
random-walk regime), and carries the initial-condition uncertainty from
the forecast origin. The forecast rows’ responses are always treated as
unknown, so passing a hindcast’s observed future values no longer leaks
them into the interval.
The AR/temporal decomposition is now tail-robust and
consistent with the reported interval. Two follow-on fixes to
the autocorrelation forecast above. (i) et_predict() built
the credible-interval / PIT / calibration draws
(predictive_draws) from only the first
n_perturb posterior draws while the variance decomposition
used all n_draws, so the interval and the decomposition
could describe different samples; the env-inflated predictive now spans
all n_draws draws (the env perturbation is recycled across
them).
temporal_var (and hence total_var) is
computed from a winsorized posterior-predictive
variance, so that when the AR coefficient’s posterior has mass at a
handful of explosive, non-stationary draws can no longer dominate the
estimate. Previously a single near-unit-root draw could inflate the raw
variance by an order of magnitude and make temporal_var
disagree wildly with the (quantile-based, tail-robust) credible interval
reported alongside it. The decomposition now tracks that interval; a
genuinely non-stationary AR posterior still yields a large temporal
share. et_predict() emits a heads-up when the raw
predictive variance greatly exceeds its robust value (the near-unit-root
regime), pointing to et_diagnose().et_predict() caps n_draws at the number
of draws in the fit (with a warning) instead of erroring, and now uses a
shared, explicit set of draw ids across posterior_predict()
and posterior_linpred() so the two are aligned
draw-for-draw.
shelf_life(): the plausible_range argument
is deprecated. Use response_scale instead.
plausible_range still works but emits a warning. It will be
removed in a future release.Optional global (Sobol) variance decomposition
(T7). New et_sobol() provides an order-independent
alternative to the additive budget: a variance-based decomposition of
the predictive mean into first-order and Sobol indices plus their
(Saltelli 2010 / Jansen estimators, with output centring for Monte Carlo
stability). The fast additive budget remains the default;
et_sobol() is for the general case where a non-linear link
or correlated drivers make the parameterdriver interaction
non-negligible, which the sequential additive split folds into its
terms.
Null-model forecast skill + shelf-life skill gate
(T6). New et_skill_score() scores a forecast
against a null model — a random walk (persistence) or climatology — with
the continuous ranked probability score (CRPS) and reports the
per-lead-time skill and the null-relative (the lead time at which the
model stops beating the null), following the ecological-forecasting
convention (Petchey et al. 2015; Wesselkamp et al. 2025; the NEON
challenge). shelf_life() gains a skill
argument: passing an et_skill_score() table gates the shelf
life so a period counts as informative only if it is both precise
skillful, preventing a precise-but-biased forecast from passing.
shelf_life() is now documented as a complement to the
accuracy-relative forecast limit, and its response_scale
now defaults to the training-response range (a documented, reproducible
default) instead of requiring an arbitrary number.
PIT / rank-histogram calibration diagnostics
(T4). New et_pit() computes the probability
integral transform of held-out observations under the reported posterior
predictive distribution (the same distribution the intervals and
et_calibrate() coverage use, environmental uncertainty
included), and et_plot_pit() renders the PIT histogram with
a uniform reference band. This is the full-distribution diagnostic — a
U-shape flags overconfident intervals, a hump flags under-confidence, a
tilt flags bias — that coverage alone cannot provide.
et_pit() supports randomized PIT for discrete predictions.
et_prediction objects now store
predictive_draws, the exact draws the interval / PIT /
calibration all share.
Pathwise driver ensembles for genuine reforecasts
(T2). et_predict() gains an
env_ensemble argument: a named list of matrices (M scenario
trajectories over the H forecast steps), one per uncertain driver
predictor. Each posterior draw is paired with a whole covariate
trajectory, so driver uncertainty is temporally coherent and
accumulates with lead time as a real driver ensemble
(CMIP members, resampled climatology) fans out — unlike the existing
env_noise, whose independent per-observation jitter
averages out and leaves the driver variance flat. This is the
recommended way to build a reforecast that uses only information
available at the issue date, and it reproduces the growing
driver-variance fraction of Dietze (2017) and Thomas et al. (2020).
env_ensemble supersedes env_noise for the
perturbation and is folded into the credible interval by
default.
shelf_life() now includes se_t_star in
the projected-mode horizon attribute. This is the delta-method standard
error of the projected crossing time t* = (τ − a) / b,
propagating the linear-fit covariance.
decompose_uncertainty() and the underlying
.decompose_from_arrays() helper now support non-Gaussian
families (Binomial, Poisson, Student-t, Negative-Binomial, Beta, Gamma)
by computing all variance components on the response scale via the
inverse link. For Gaussian identity the result is numerically
unchanged.
et_predict() gains an n_env_draws
argument. Setting it > 1 averages multiple independent perturbations
per posterior draw, reducing Monte Carlo noise in the environmental
variance estimate. The decomposition data frame gains a
v_env_mcse column reporting the chi-squared standard error
of env_var.
decompose_uncertainty() now reports a fourth
temporal_var component when the model formula contains an
autocorrelation term (ar(), ma(),
arma(), cosy(), unstr(),
sar(), or car()). The component is computed as
pmax(0, total_var - (param_var + env_var + residual_var))
and captures the autocorrelation-induced predictive variance beyond the
iid sum, so the four components reconstruct total_var
modulo Monte Carlo error. residual_var for autocor models
is interpreted as the innovation variance (not the stationary marginal
variance). et_plot_decomposition() adds a Temporal segment
to the stacked bars automatically when the column is present, and
print.et_prediction() includes the new row in its
summary.
et_plot_calibration() previously only recognised a
column literally named group as the sub-group identifier.
Calibration data frames built by binding per-group results with
descriptive column names (e.g. cluster_id,
species, regime) were silently collapsed into
a single un-grouped series, producing plots with multiple overlapping
points per nominal level and a zig-zagging connecting line. The function
now auto-detects any single non-canonical column (anything other than
ci_level, nominal,
observed_coverage, n_obs,
calibration_error, sharpness) with more than
one unique value and uses it as the grouping variable. A new
group_col argument allows the grouping column to be set
explicitly, or group_col = NA to force a single un-grouped
series.Initial CRAN submission.
Initial CRAN release.
Added full Bayesian error propagation pipeline using
brms.
Added three-way variance decomposition and forecast shelf life metrics.