Methods for forecasting hierarchical or grouped time series.
# S3 method for gts
forecast(
object,
h = ifelse(frequency(object$bts) > 1L, 2L * frequency(object$bts), 10L),
method = c("comb", "bu", "mo", "tdgsa", "tdgsf", "tdfp"),
weights = c("wls", "ols", "mint", "nseries"),
fmethod = c("ets", "arima", "rw"),
algorithms = c("lu", "cg", "chol", "recursive", "slm"),
covariance = c("shr", "sam"),
nonnegative = FALSE,
control.nn = list(),
keep.fitted = FALSE,
keep.resid = FALSE,
positive = FALSE,
lambda = NULL,
level,
FUN = NULL,
xreg = NULL,
newxreg = NULL,
parallel = FALSE,
num.cores = 2,
...
)A forecasted hierarchical/grouped time series of class gts.
Hierarchical or grouped time series object of class
{gts}
Forecast horizon
Method for distributing forecasts within the hierarchy. See details
Weights used for "optimal combination" method:
weights="ols" uses an unweighted combination (as described in Hyndman
et al 2011); weights="wls" uses weights based on forecast variances
(as described in Hyndman et al 2016); weights="mint" uses a full
covariance estimate to determine the weights (as described in Wickramasuriya et al
2019); weights="nseries" uses weights based on the number of series
aggregated at each node.
Forecasting method to use for each series.
An algorithm to be used for computing the combination
forecasts (when method=="comb"). The combination forecasts are based
on an ill-conditioned regression model. "lu" indicates LU decomposition is
used; "cg" indicates a conjugate gradient method; "chol" corresponds to a
Cholesky decomposition; "recursive" indicates the recursive hierarchical
algorithm of Hyndman et al (2016); "slm" uses sparse linear regression. Note
that algorithms = "recursive" and algorithms = "slm" cannot be
used if weights="mint".
Type of the covariance matrix to be used with
weights="mint": either a shrinkage estimator ("shr") with
shrinkage towards the diagonal; or a sample covariance matrix
("sam").
Logical. Should the reconciled forecasts be non-negative?
A list of control parameters to be passed on to the block principal pivoting algorithm. See 'Details'.
If TRUE, keep fitted values at the bottom level.
If TRUE, keep residuals at the bottom level.
If TRUE, forecasts are forced to be strictly positive (by
setting lambda=0).
Box-Cox transformation parameter.
Level used for "middle-out" method (only used when method
= "mo").
A user-defined function that returns an object which can be
passed to the forecast function. It is applied to all series in order
to generate base forecasts. When FUN is not NULL,
fmethod, positive and lambda are all ignored. Suitable
values for FUN are tbats and
stlf for example.
When fmethod = "arima", a vector or matrix of external
regressors used for modelling, which must have the same number of rows as
the original univariate time series
When fmethod = "arima", a vector or matrix of external
regressors used for forecasting, which must have the same number of rows as
the h forecast horizon
If TRUE, import parallel package to allow parallel
processing.
If parallel = TRUE, specify how many cores are going to be
used.
Other arguments passed to ets,
auto.arima or FUN.
Earo Wang, Rob J Hyndman and Shanika L Wickramasuriya
Base methods implemented include ETS, ARIMA and the naive (random walk) models. Forecasts are distributed in the hierarchy using bottom-up, top-down, middle-out and optimal combination methods.
Three top-down methods are available: the two Gross-Sohl methods and the
forecast-proportion approach of Hyndman, Ahmed, and Athanasopoulos (2011).
The "middle-out" method "mo" uses bottom-up ("bu") for levels
higher than level and top-down forecast proportions ("tdfp")
for levels lower than level.
For non-hierarchical grouped data, only bottom-up and combination methods are possible, as any method involving top-down disaggregation requires a hierarchical ordering of groups.
When xreg and newxreg are passed, the same covariates are
applied to every series in the hierarchy.
The control.nn argument is a list that can supply any of the following components:
ptypePermutation method to be used: "fixed" or "random". Defaults to "fixed".
parThe number of full exchange rules that may be tried. Defaults to 10.
gtolThe tolerance of the convergence criteria. Defaults to sqrt(.Machine$double.eps).
Athanasopoulos, G., Ahmed, R. A., & Hyndman, R. J. (2009). Hierarchical forecasts for Australian domestic tourism, International Journal of Forecasting, 25, 146-166.
Hyndman, R. J., Ahmed, R. A., Athanasopoulos, G., & Shang, H. L. (2011). Optimal combination forecasts for hierarchical time series. Computational Statistics and Data Analysis, 55(9), 2579--2589. https://robjhyndman.com/publications/hierarchical/
Hyndman, R. J., Lee, A., & Wang, E. (2016). Fast computation of reconciled forecasts for hierarchical and grouped time series. Computational Statistics and Data Analysis, 97, 16--32. https://robjhyndman.com/publications/hgts/
Wickramasuriya, S. L., Athanasopoulos, G., & Hyndman, R. J. (2019). Optimal forecast reconciliation for hierarchical and grouped time series through trace minimization. Journal of the American Statistical Association, 114(526), 804--819. https://robjhyndman.com/publications/mint/
Wickramasuriya, S. L., Turlach, B. A., & Hyndman, R. J. (to appear). Optimal non-negative forecast reconciliation. Statistics and Computing. https://robjhyndman.com/publications/nnmint/
Gross, C., & Sohl, J. (1990). Dissagregation methods to expedite product line forecasting, Journal of Forecasting, 9, 233--254.
hts, gts,
plot.gts, accuracy.gts
forecast(htseg1, h = 10, method = "bu", fmethod = "arima")
if (FALSE) {
forecast(
htseg2, h = 10, method = "comb", algorithms = "lu",
FUN = function(x) tbats(x, use.parallel = FALSE)
)
}
Run the code above in your browser using DataLab