bpv non-negative forecast reconciliation
algorithm (experimental) based on the cross-sectional framework
presented in Wickramasuriya et al. (2020) and, now, extended for the
temporal and cross-temporal framework:
oasd cross-sectional covariance matrix
(experimental), implementing an oracle shrunk covariance estimation
(Ando and Xiao, 2023):
bounds parameter to enable bounded forecast
reconciliation for csrec(), terec(), and
ctrec() functions;set_bounds() function to define custom
bounds for reconciliation.agg_order factors was
selected in iterec(), tcsrec(), and
cstrec() functions.Note – The latest release of FoReco introduces significant changes to its function notation and adds several new features. This major update, FoReco 1.0, is not compatible with previous versions due to the substantial changes made to the package’s core structure. The previous version is available on Github (docs).
Due to the significant changes in FoReco 1.0, users are advised to carefully review the updated documentation and examples before using the new version. The latest documentation and release notes are available on danigiro.github.io/FoReco/
Updated Function Notation: All functions related
to cross-sectional, temporal, and cross-temporal frameworks now use the
prefixes cs, te, and
ct, respectively. For example, the optimal combination
reconciliation functions are now csrec(),
terec(), and ctrec().
Simplified Function Outputs: Reconciliation
functions now return only matrices. Additional information can be
accessed using attr(., "FoReco") or the
recoinfo() function.
New Datasets: Two new datasets,
itagdp (Italian Quarterly National Accounts) and
vndata (Australian Tourism Demand), are included along with
their respective aggregation or constraint matrices.
Classic Approach: The middle-out approach
(csmo(), temo(), and ctmo()) has
been implemented alongside the classic bottom-up (csbu(),
tebu(), and ctbu()) and top-down
(cstd(), tetd(), and cttd())
methods.
Level Conditional Coherent Reconciliation: Level
conditional coherent reconciliation is now available for all
constraints: cslcc() (cross-sectional),
telcc() (temporal), and ctlcc()
(cross-temporal).
Immutable reconciliation: The
immutable() parameter has been added to the reconciliation
functions (csrec(), terec(), and
ctrec()) to prevent the base forecasts from being modified
with both the structural (approach='strc') and the
zero-constrained (approach='proj') approach.
Balanced and unbalanced hierarchy: added
balance_hierarchy() and unbalance_hierarchy()
for dealing with balanced and unbalanced hierarchies.
Projection Matrix Functions: Functions
csprojmat(), teprojmat(), and
ctprojmat() have been added to obtain projection
matrices.
Covariance Matrix Functions: Functions
cscov(), tecov(), and ctcov()
have been added to obtain covariance matrices.
Function Renaming: Several functions have been renamed to improve consistency and clarity
Cmatrix() -> df2aggmat()hts_tools() -> cstools()thf_tools() -> tetools()ctf_tools() -> cttools()agg_ts() -> aggts()residuals_matrix() -> res2matrix()boot_cs() -> csboot()boot_te() -> teboot()boot_ct() -> ctboot()htsrec() -> csrec()thfrec() -> terec()octrec() -> ctrec()lccrec() -> cslcc()boot_cs(), boot_te() and
boot_ct() to draw samples from, respectively,
cross-sectional, temporal and cross-temporal joint (block)
bootstrap.ctbu() inputs;FoReco2matrix() to transform FoReco forecasts
input and output in a list of matrix/vector class;agg_ts(): non-overlapping temporal aggregation of
a time series according to a specific aggregation order;arrange_hres() and
residuals_matrix() functions to arrange residuals for the
covariance matrix under Gaussianity.octrec();lcmat() function.htsrec();score_index();bounds param when type = "S" in
htsrec(), thfrec() and
octrec();v
param in htsrec(), thfrec() and
octrec() - experimental;ut2c() and srref().lccrec() (now the function
returns the Level Conditional Coherent or the Combined Conditional
Coherent forecasts);keep = "list".htsrec(),
thfrec() and octrec() to introduce a list of
h covariance matrices in the parameters W and
Omega, where h stands for the forecast horizon
(note that for thfrec() and octrec() this is
the forecast horizon of the entire cycle);Sstruc is no more avaible in
octrec() and ctf_tools(). FoReco uses a fast
algorithm to compute Scheck, so no external input is
needed;ctf_tools() (added
Ccheck, Htcheck, Scheck, removed
Cstruc, Sstruc), hts_tools()
(added C) and thf_tools() (added
m);nn_type) in htsrec(),
thfrec() and octrec();tdrec();levrec() (cross-sectional, temporal and
cross-temporal).octrec() it is also possible to introduce the
Ω covariance matrix variant through the
Omega parameter and not only the W variant
with the W parameter;tcsrec(), cstrec() and
iterec(). In the iterec() function the
maxit parameter has been replaced by itmax,
however for the moment maxit is still supported;iterec() (norm
parameter).bounds param in htsrec(),
thfrec() and octrec();oct_bounds() to organize the bounds on a
specific dimension (i.e. only cross-sectional or only temporal) in a
cross-temporal framework;ut2c() and srref() to develop a
cross-sectional structural representation starting from a zero
constraints kernel matrix;score_index() the calculation of multiple
forecast horizons index (like 1:6) and multiple cross-sectional levels
for a forecasting experiment.Minore release, fixing some bugs and the documentation
iterec() when calculating the
incoherenceAverage relative accuracy indices”