NEWS.md

ForeCA Release Notes

0.2.7

• add pspectrum from psd package as a replacement for sapa::SDF. Not recommended for use in production yet since psd::psectrum does not preserve spectrum estimates from multivariate spectrum to univariate, i.e., spectrum of linear combination does not equal linear combination (quadratic form) of spectrum. psd::pspectrum an be used to compute univariate forecastability estimates based on a better (smoothed) estimates of periodogram compared to raw spec.pgram.

v0.2.6-1

• remove all dependencies and calls to sapa R package. In particular this means no "wosa", "multitaper", or "direct" spectral method argument anymore. Instead rely on "mvspec" in the astsa package.

  This is (unfortunately) necessary since ifultools and all its dependencies were removed from CRAN by 2020-06-09. sapa was the workhorse package for ForeCA.

v0.2.5 (May 2020)

• make ForeCA compatible with R 4.0.0
• improve test coverage and error messages (move away from expect_true wherever possible)

v0.2.4 (April 2016)

New

• added an introductory R markdown vignette using knitr

Bug fixes

• quadratic_form now computes the vector product with the transpose and conjugate of the vector, i.e., the Hermitian of the vector. Also added small speed up using crossprod function.
• sqrt_matrix did not work correctly for singular matrices (zero eigenvalue). Fixed issue with numerical precision errors testing for equality to zero. Now throws an error if matrix is not of full rank and the inverse is required.

v0.2.2 (April 2015)

New

• fill_hermitian() is now a public function (and added tests)
• added foreca.EM.E_and_M_step as a wrapper for foreca.EM.E_step followed by foreca.EM.M_step.
• add a TRUE/FALSE "normalized" attribute to objects of class mvspectrum. This speeds up computation and checks significantly since check_mvspectrum_normalized only needs to check the attribute, rather than computing the sum and comparing it to identity matrix.
• removed "lag window" option for mvspectrum since it does not work well with normalize_mvspectrum.
• added a plot.mvspectrum S3 method

Bug fixes

• normalize_mvspectrum() normalizes by left and right multiplication of inverse square root of frequency total aggregate estimate. This fixes the issue with different results of normalization based on multivariate vs. univariate spectra estimation.
• removed a print() statement when running foreca() (displayed Omega scores of the final ForeCs)
• sqrt_matrix threw an error if input matrix had complex-valued eigen-values (since it checked for negative values). Fixed now.
• specify nrow in diag() if n.comp=1 in foreca.multiple_weightvectors (otherwise it generated non-comformable arrays error).

Minor changes / improvements

• fixed real/imaginary problem with the eigenvalues of Hermitian matrices in test_mvspectrum()
• change tolerance level in EM convergence to 1e-6 (see complete_algorithm_control())

v0.2.0 (Jan 2015)

A bug-fix release and more modular, less repetitive coding under the hood; results in improved performance of the main algorithms.

Main notable change for users: to specify spectrum and entropy estimation use spectrum.control and entropy.control. Otherwise no relevant visible user-interface changes.

Also review the manual; it has been very thoroughly reviewed.

New

• a couple of new functions for more modular code and easier testing:
  • sqrt_matrix
  • complete_entropy_control
  • complete_spectrum_control
  • complete_algorithm_control
  • check_whitened
  • check_mvspectrum_normalized
  • weightvector2entropy_wcov
  • mvpgram
• a print.foreca S3 method
• Omega and several foreca.* functions changed arguments:
  • spectrum.conrol = list(...) is the new way to specify spectrum.method, smoothing.
  • entropy.control = list(...) is the new way to specify entropy.method, threshold, prior.weight.
• made "mvfft" default for spectrum estimation, using mvfft in R. This avoids the requirement to have sapa or astsa installed.
• general foreca.one_weightvector and foreca.multiple_weightvectors wrappers for more modular and less repititve coding.
• By default entropy is smoothed by a mixture with a uniform prior distribution with prior.weight = 1e-03 in order to avoid log(0) throughout the computations. If you don't want this, you have to explicitly specify prior.weight = 0 in the entropy.control argument.
• formally added tests using the testthat package
• added list of all available spectrum.methods for estimation of the spectrum/spectral density in the complete_spectrum_control help page.
• all foreca.*() functions only accept whitened time series U, not series. Only foreca() directly accepts the original (unwhitened) series. Use U <- whiten(series)$U to obtain the whitened series.

Removed

• foreca.EM: use foreca(..., algorithm.type = "EM") directly.
• dependency on R packages:
  • R.utils: R.utils::wrap got replaced by base::aperm.
  • sapa and astsa: in its basic form the ForeCA package now runs in base R! However, it is still highly recommended to use either the sapa or astsa package (users can use their algorihms via the method argument in several functions.)
• spectral_entropy now only works with a spectral density input f.U; not with series or spectrum.estimate.

Bug fixes

• whiten() (thanks to Bjoern Weghenkel for pointing this out):
  • forgot to center U; if data was centered to start with, then this didn't affect results.
  • fixed dewhitening and whitening matrices
• initialize_weightvector used the minimum Omega vector, not maximum, for method = "max".

Minor changes/improvements

• thourough revision of documentation:
  • better explanations
  • fixed typos or code / documentation mismatch
  • more concise (hopefully)
• mvspectrum: the smooth univariate spectrum estimate (smoothing = TRUE) uses an exponential distribution with a logarithmic link function in the mgcv::gam function.
• discrete_entropy:
  • The uniform distribution is now added using a mixture type model, rather than an absolute addition to original distribution and then renormalizing. Thus the prior.weight is now between 0 and 1, not just greater than 0. can be done just with the prior.weight argument.
  • Only ''MLE'' is a valid method now. If you want to use ''smoothed'', then this
• lower and upper limit arguments for continuous_entropy changed from a and b to lower and upper.
• eigen() calls got symmetric = TRUE. About 3x faster.
• The random methods in initialize_weightvector have the r prefix, i.e., "rnorm" instead of "norm". Same for "cauchy" and "unif".
• foreca.EM.opt_weightvector changed to foreca.EM.one_weightvector.
• the trace plots of the weights are always smooth curves (and don't jump betweeen +/- 1 times the weightvector).

---

v0.1

• changed underscore (_) in argument names to dot (.). E.g., max_iter to max.iter
• cleaned up code for better readability
  • changed = to <- assignment operator (big thanks to tidy.source())
  • removed unnecessary code (thanks checkUsage())
• cleaned up NEWS file and edited to conform to proper markdown format
• changed method argument in foreca() to algorithm.type
• moved function to initialize a weightvector from the EM class, to its own function (initialize_weightvector())
• replaced AR spectrum estimation to "burg" in spec.ar()
• moved tol, nstart, and max.iter in foreCA.EM() into a control list (where nstart became num.starts)
• improved display of plot.foreca.EM.opt_weightvector

Bug fixes

• fix bug in mvspectrum2wcov()
• fix bug in mvspectrum() (set detrend = FALSE and fast = FALSE in astsa::mvspec)
• make all spectral estimate functions return the same number of frequencies (T/2 +/- 1 depending on even/odd sample size)
• fix bug in fill_symmetric() (double counting diagonal; only affected SDF type estimation)

v0.0.9

• changed capitalized ForeCA names to lowercase (except for abbreviations such as EM or MLE). E.g., ForeCA.EM to foreca.EM;
• changed foreca.one_weightvector() to foreca.EM.opt_weightvector()

v0.0.8.1

• added many additional functions in the package, including the main algorithm ForeCA.EM
• changed to Roxygen2 documentation

v0.0.1 (initial release)

• base functions to estimate (spectral) entropies and Omega for (multivariate) time series
• first draft of documentation
• simple examples in help files

First version 0.0.1 written by Georg M. Goerg on May 14, 2012.

Bugs & feature requests

• check what is the best way to store the 3D array spectrum
  • frequency last or first?
  • data.table package?
• include more spectrum estimators
• include a continuous estimator of spectral entropy; can be used, e.g., to compute entropy for a fitted AR spectrum
• compressed sensing type of sparsity in the spectrum; currently only by a heuristic thresholding rule
• nice plotting for 3D spectra?