Grouped Time-Series Forecasting

This learner supports prediction using grouped time-series modeling, using hts. Fitting is done with hts and prediction is performed via forecast.gts.

Format

R6Class object.

Value

Learner object with methods for training and prediction. See Lrnr_base for documentation on learners.

Parameters

method): Method for distributing forecasts within hierarchy. See details of forecast.gts.
weights): 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 2015); weights="mint" uses a full covariance estimate to determine the weights (as described in Hyndman et al 2016); weights="nseries" uses weights based on the number of series aggregated at each node.
fmethod): Forecasting method to use for each series.
algorithms): 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 (2015); "slm" uses sparse linear regression. Note that algorithms = "recursive" and algorithms = "slm" cannot be used if weights="mint".
covariance): 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").
keep.fitted): If TRUE, keep fitted values at the bottom level.
keep.resid): If TRUE, keep residuals at the bottom level.
positive): If TRUE, forecasts are forced to be strictly positive (by setting lambda=0).
lambda): Box-Cox transformation parameter.
level: Level used for "middle-out" method (only used when method = "mo").
parallel: If TRUE, import parallel to allow parallel processing.
num.cores: If parallel = TRUE, specify how many cores are going to be used.

Examples

# Example adapted from hts package manual
# The hierarchical structure looks like 2 child nodes associated with level 1,
# which are followed by 3 and 2 sub-child nodes respectively at level 2.
library(hts)
#> Loading required package: forecast

set.seed(3274)
abc <- as.data.table(5 + matrix(sort(rnorm(200)), ncol = 4, nrow = 50))
setnames(abc, paste("Series", 1:ncol(abc), sep = "_"))
abc[, time := .I]
#>     Series_1 Series_2 Series_3 Series_4 time
#>  1: 2.607712 4.453525 5.059647 5.674443    1
#>  2: 2.770780 4.468715 5.081054 5.686761    2
#>  3: 2.782223 4.482786 5.102684 5.707513    3
#>  4: 2.787655 4.483375 5.115142 5.739496    4
#>  5: 2.932073 4.489276 5.136250 5.748407    5
#>  6: 3.064115 4.527663 5.141674 5.789793    6
#>  7: 3.123294 4.549361 5.144565 5.794407    7
#>  8: 3.249828 4.549487 5.158862 5.799323    8
#>  9: 3.370915 4.558805 5.161763 5.799465    9
#> 10: 3.379109 4.560854 5.198417 5.838728   10
#> 11: 3.383535 4.562303 5.199713 5.861504   11
#> 12: 3.420078 4.562310 5.212496 5.865641   12
#> 13: 3.425742 4.571433 5.215093 5.869555   13
#> 14: 3.495619 4.589284 5.216852 5.909086   14
#> 15: 3.497682 4.594580 5.217866 5.947755   15
#> 16: 3.502086 4.617152 5.228778 5.963670   16
#> 17: 3.507000 4.620320 5.240832 5.994254   17
#> 18: 3.539549 4.629534 5.257752 6.012084   18
#> 19: 3.558871 4.638255 5.304387 6.021723   19
#> 20: 3.654411 4.656876 5.304651 6.073125   20
#> 21: 3.691043 4.657565 5.306511 6.082792   21
#> 22: 3.708930 4.662122 5.367066 6.118491   22
#> 23: 3.710579 4.665984 5.381140 6.133808   23
#> 24: 3.719483 4.696833 5.388976 6.144857   24
#> 25: 3.734375 4.698856 5.390283 6.166144   25
#> 26: 3.743854 4.704998 5.397072 6.198817   26
#> 27: 3.881388 4.714980 5.458463 6.228678   27
#> 28: 3.900640 4.735690 5.460558 6.248769   28
#> 29: 3.908360 4.739398 5.464411 6.251769   29
#> 30: 3.952202 4.808120 5.468979 6.289440   30
#> 31: 3.970051 4.826650 5.496115 6.302676   31
#> 32: 4.055186 4.831577 5.513275 6.379657   32
#> 33: 4.096847 4.867235 5.519347 6.418440   33
#> 34: 4.108913 4.870614 5.523854 6.418850   34
#> 35: 4.109187 4.886651 5.525574 6.431287   35
#> 36: 4.138922 4.888021 5.546059 6.502432   36
#> 37: 4.145324 4.897311 5.548523 6.558697   37
#> 38: 4.149141 4.898081 5.556770 6.559662   38
#> 39: 4.179869 4.901912 5.563149 6.581637   39
#> 40: 4.250431 4.905990 5.566242 6.612510   40
#> 41: 4.306163 4.920350 5.576517 6.632175   41
#> 42: 4.315984 4.931794 5.576623 6.674475   42
#> 43: 4.343608 4.935432 5.588910 6.728559   43
#> 44: 4.347261 4.945405 5.622140 6.796547   44
#> 45: 4.357179 4.965257 5.657682 6.822929   45
#> 46: 4.359392 4.982516 5.659091 6.955815   46
#> 47: 4.368390 4.990881 5.659724 7.041288   47
#> 48: 4.394952 5.026946 5.662659 7.406428   48
#> 49: 4.398161 5.031637 5.669859 7.551534   49
#> 50: 4.412912 5.043044 5.671561 7.819792   50
#>     Series_1 Series_2 Series_3 Series_4 time
grps <- rbind(c(1, 1, 2, 2), c(1, 2, 1, 2))
horizon <- 12
suppressWarnings(abc_long <- melt(abc, id = "time", variable.name = "series"))

# create sl3 task (no outcome for hierarchical/grouped series)
node_list <- list(outcome = "value", time = "time", id = "series")
train_task <- sl3_Task$new(data = abc_long, nodes = node_list)
test_data <- expand.grid(time = 51:55, series = unique(abc_long$series))
test_data <- as.data.table(test_data)[, value := 0]
test_task <- sl3_Task$new(data = test_data, nodes = node_list)

gts_learner <- Lrnr_gts$new()
gts_learner_fit <- gts_learner$train(train_task)
gts_learner_preds <- gts_learner_fit$predict(test_task)

Format

Value

Parameters

See also

Examples