NEWS
ivx 2.0.0
Breaking changes
summary() coefficient tables now report Std. Error and t value next to the
individual Wald statistics; the fitted object stores se and tstat.
The table gains two columns, so code indexing coef(summary(x)) by
position must use column names ("Wald Ind", "Pr(> chi)") instead.
Deprecated
drop1(), add1() and step() on ivx objects. They compare
residual sums of squares of an IV estimator, which is not a valid selection
criterion; use the IVX Wald tests in summary().
New estimators and tests
- New
ivx_ra() / ivx_ra_fit(): the residual-augmented (bias-reduced) IVX
estimator of Demetrescu & Rodrigues (2022) with its heteroskedasticity-robust
standard errors; returns an ivx object so summary(), vcov() etc. apply.
ivx_ra() gains horizon: for horizon > 1 it is the transformed-regression
long-horizon test of Demetrescu, Rodrigues & Taylor (2023), which handles the
overlap of the long-horizon regression without HAC estimation.
- New
ivx_qr() / ivx_qr_fit(): the IVX-QR quantile predictability test of
Lee (2016, Proposition 3.2) via quantreg (in Suggests); returns the
estimated QR endogeneity rho_tau for the paper's tuning rule.
- New
ivx_qr_boot(): moving block bootstrap percentile intervals and
p-values for IVX-QR (Fan & Lee, 2019), robust to conditional
heteroskedasticity and to the sparsity estimate.
- New
ivx_episodic(): subsample (rolling, forward and backward recursive)
IVX tests for pockets of predictability with sup/inf functionals and wild
bootstrap p-values (Demetrescu et al. 2022, 2023 Section 3.2).
- New
ivx_sys() / ivx_sys_fit(): systems of predictive regressions with a
matrix response (cbind(y1, y2) ~ x), short and long horizon, with the
Kronecker-form IVX-Wald covariance of Kostakis et al. (2023); reports joint,
per-equation and individual Wald statistics.
- New
ivx_boot() implementing the residual wild bootstrap and fixed regressor
wild bootstrap of Demetrescu et al. (2023), returning bootstrap p-values for
the joint and individual Wald statistics and one-sided t-tests. Supports
cores > 1 via the parallel package; the regressor recursion of the
residual wild bootstrap runs in C++.
- New
ivx_iv() / ivx_iv_fit(): the 2SLS predictability tests of Breitung &
Demetrescu (2015) with fractional-difference, long-difference and sine
instruments and their recommended combination (IVcomb), Eicker-White
standard errors.
- New
arm() / arm_fit(): the multipredictor augmented regression method of
Amihud, Hurvich & Wang (2009) - reduced-bias OLS with Nicholls-Pope corrected
VAR(1) residuals as control variables and the paper's covariance estimator;
a non-IVX benchmark for stationary persistent predictors.
- New
hlt_test(): the hybrid switching t-test of Harvey, Leybourne & Taylor
(2021) - standard or quasi-GLS-demeaned t-ratio with the paper's conservative
critical values under strong persistence, normal critical values under weak
persistence (ADF/MBIC switch).
- New
el_test(): the unified empirical likelihood test of Liu, Yang, Cai &
Peng (2019) for the predictive regression augmented with the lagged
difference of the predictor; chi-square profile EL ratios whatever the
persistence of the predictor, no tuning parameters.
- New
cy_test(): the Bonferroni Q-test of Campbell & Yogo (2006) (the feasible
Cavanagh, Elliott & Stock 1995 approach): DF-GLS confidence interval for the
largest root inverted from simulated local-to-unity quantiles, Table 2 levels,
Q-estimates with the AR(p) correction of Appendix A.
- New
elliott_cf(): the control-function predictive regression of Elliott
(2011) with user-supplied orthogonalising covariates and their lags, Wald
test with Eicker-White standard errors and the remaining innovation
correlation as a diagnostic.
New arguments and methods
ivx() and ivx_fit() gain beta, cz and bandwidth arguments that expose
the IVX instrument tuning (previously hard-coded to the Kostakis et al. (2015)
values) and the Newey-West bandwidth.
ivx() gains robust = TRUE for Eicker-White (heteroskedasticity-robust) IVX
standard errors (Demetrescu, Georgiev, Rodrigues & Taylor, 2023).
ivx() and ivx_fit() gain lag_y = TRUE: the lag-augmented IVX regression
of Demetrescu (2014), which adds the lagged dependent variable (instrumented
by itself) to raise local power under strong persistence and endogeneity.
The joint Wald statistic tests the predictors only.
ivx_ar() and ivx_ar_fit() accept the same beta, cz, bandwidth and
robust arguments as ivx().
- New
predict() method for ivx() fits: the one-step-ahead conditional mean
for the predictors in newdata, so coefficients estimated on a training
sample can be projected out of sample (e.g., a fundamental price path).
- New
nobs() method for ivx objects (and the extensions inheriting the class).
extract() (texreg) now reports IVX standard errors instead of Wald statistics
in the se slot.
Bug fixes and output changes
- Fixed:
ivx_ar(ar = "auto") ignored ar_ic and always chose the AR order
by AIC; ar_ic = "bic" (the default) now selects by BIC. ar = "forecast"
stops with an informative error when forecast is not installed.
- Arguments aligned with the papers they implement:
ivx_iv() d is now the
memory of the fractional instrument (1 - L)^(1 - d) x[t-1] and the long
difference is x[t-1] - x[t-k] with k <= t - 1, as in
Breitung & Demetrescu (2015, Section 2.2; the default d = 1/2 is
unchanged); ivx_episodic() window is the shortest window for every
scheme and defaults to the settings of Demetrescu et al. (2022): 1/3 rolling,
1/4 recursive; the ivx_boot() / ivx_episodic() ar_max default is the
range of Demetrescu et al. (2023), floor(4 * (T / 100)^(1/4)) + 1;
elliott_cf() defaults to the paper's conventional Wald test
(robust = FALSE).
- Fixed: weighted fits ignored the
horizon argument; zero weights made
ivx() fail (the dropped observations now get NA residuals and fitted
values, with the same coefficients as fitting on the kept rows).
- Fixed:
ac_test_bg() dispatched to the Box-Pierce method; ac_test_lb() and
ac_test_bp() reported wrong p-values for non-consecutive lag vectors;
the Breusch-Godfrey result now has the same ac_test_ class and pval
attribute as the other tests. case.names() no longer returns an empty vector.
print() of ac_test() on a numeric vector no longer errors (it indexed a
Breusch-Godfrey column that only the ivx method computes).
ac_test() and ac_test_*() now stop with an informative error on ivx_ar,
ivx_ra, ivx_qr, ivx_iv and arm fits, which do not store the OLS
residuals the tests use; previously they failed with an obscure error.
summary() of ivx_ar() printed Auto () without the information
criterion; it now prints e.g. Auto (BIC), and ar = "forecast" fits are
labelled forecast (BIC) instead of Fixed.
print() of hlt_test() and cy_test() splits its output into short lines
(at most 70 characters).
- All formula interfaces share one model-frame routine, so
- 1 in a formula
warns consistently and a matrix response is rejected consistently.
- The
ylpc help page cited the wrong title, issue and pages for its source;
it now cites Yang, Long, Peng & Cai (2020), JASA 115(532), 1598-1619.
Documentation
- New vignette "Choosing a test: a decision workflow" (
vignette("workflow"))
that walks from the plain ivx() fit through the diagnostics (ac_test(),
delta(), ARCH check) to the extension each one calls for.
- New vignette "Rolling IVX tests for bubble detection" showing how to build the
rolling-window IVX test of Pavlidis, Paya & Peel (2017) from
ivx() (#2).
- New vignettes for each methodology (
ivx, ivx-sys, ivx-ar, ivx-ra,
ivx-qr, robust-inference, ivx-episodic) with the underlying
statistics, replication results and caveats; pkgdown site reorganised
(Bootstrap 5, MathJax rendering, grouped reference and articles).
- Documented that the long-horizon statistic (
horizon > 1) is the modified
IVX-Wald of Kostakis, Magdalinos & Stamatogiannis (2023), eqs (15)/(23),
after auditing the implementation against the paper.
ivx 1.1.1 (2025-09-20)
- Fixed
ivx_ar() when auto.arima() selects the AR order (#3).
- Added an S3 method for
ac_test_bg(), and extract() for ivx now checks
that texreg is installed.
- Removed the C++11 specification from
src/Makevars and src/Makevars.win
to address a CRAN NOTE; inst/CITATION now uses bibentry() instead of the
old-style citEntry().
- Documentation and pkgdown fixes.
ivx 1.1.0 (2020-11-24)
- Added
ivx_ar(), which implements the instrumental variable based Wald
statistic of Yang, B., Long, W., Peng, L., & Cai, Z. (2020). It accounts for
serial correlation and heteroscedasticity in the error terms of the linear
predictive regression model.
- Added the Yang et al. (2020) dataset named
ylpc.
- Breaking: renamed the
monthly and quarterly datasets to kms and
kms_quarterly.
- Removed the dependency on
tibble and magrittr.
- Added
texreg support that converts regression output to LaTeX or HTML
tables: extract() methods for ivx and ivx_ar extract coefficients and
GOF measures from the fitted object.
ivx 1.0.0 (2019-05-04)
- Added a
NEWS.md file to track changes to the package.