model_backward: Nonparametric Additive Model with Backward Elimination

In smimodel: Sparse Multiple Index Models for Nonparametric Forecasting

Nonparametric Additive Model with Backward Elimination

Description

Fits a nonparametric additive model, with simultaneous variable selection through a backward elimination procedure as proposed by Fan and Hyndman (2012).

Usage

model_backward(
  data,
  val.data,
  yvar,
  neighbour = 0,
  family = gaussian(),
  s.vars = NULL,
  s.basedim = NULL,
  linear.vars = NULL,
  refit = TRUE,
  tol = 0.001,
  parallel = FALSE,
  workers = NULL,
  exclude.trunc = NULL,
  recursive = FALSE,
  recursive_colRange = NULL,
  verbose = FALSE
)

Arguments

data	Training data set on which models will be trained. Must be a data set of class tsibble.(Make sure there are no additional date or time related variables except for the index of the tsibble). If multiple models are fitted, the grouping variable should be the key of the tsibble. If a key is not specified, a dummy key with only one level will be created.
val.data	Validation data set. (The data set on which the model selection will be performed.) Must be a data set of class tsibble.
yvar	Name of the response variable as a character string.
neighbour	If multiple models are fitted: Number of neighbours of each key (i.e. grouping variable) to be considered in model fitting to handle smoothing over the key. Should be an integer. If neighbour = x, x number of keys before the key of interest and x number of keys after the key of interest are grouped together for model fitting. The default is neighbour = 0 (i.e. no neighbours are considered for model fitting).
family	A description of the error distribution and link function to be used in the model (see glm and family).
s.vars	A character vector of names of the predictor variables for which splines should be fitted (i.e. non-linear predictors).
s.basedim	Dimension of the bases used to represent the smooth terms corresponding to s.vars. (For more information refer mgcv::s().)
linear.vars	A character vector of names of the predictor variables that should be included linearly into the model (i.e. linear predictors).
refit	Whether to refit the model combining training and validation sets after model selection. If FALSE, the final model will be estimated only on the training set.
tol	Tolerance for the ratio of relative change in validation set MSE, used in model selection.
parallel	Whether to use parallel computing in model selection or not.
workers	If parallel = TRUE, number of workers to use.
exclude.trunc	The names of the predictor variables that should not be truncated for stable predictions as a character string. (Since the nonlinear functions are estimated using splines, extrapolation is not desirable. Hence, if any predictor variable in val.data that is treated non-linearly in the estimated model, will be truncated to be in the in-sample range before obtaining predictions. If any variables are listed here will be excluded from such truncation.)
recursive	Whether to obtain recursive forecasts or not (default - FALSE).
recursive_colRange	If recursive = TRUE, the range of column numbers in val.data to be filled with forecasts. Recursive/autoregressive forecasting is required when the lags of the response variable itself are used as predictor variables into the model. Make sure such lagged variables are positioned together in increasing lag order (i.e. lag_1, lag_2, ..., lag_m, lag_m = maximum lag used) in val.data, with no break in the lagged variable sequence even if some of the intermediate lags are not used as predictors.
verbose	Logical; controls whether progress messages (model indices) are printed during fitting. Defaults to FALSE.

Details

This function fits a nonparametric additive model formulated through Backward Elimination, as proposed by Fan and Hyndman (2012). The process starts with all predictors included in an additive model, and predictors are progressively omitted until the best model is obtained based on the validation set. Once the best model is obtained, the final model is re-fitted for the data set combining training and validation sets. For more details see reference.

Value

An object of class backward. This is a tibble with two columns:

key	The level of the grouping variable (i.e. key of the training data set).
fit	Information of the fitted model corresponding to the key.

Each row of the column fit is an object of class gam. For details refer mgcv::gamObject.

References

Fan, S. & Hyndman, R.J. (2012). Short-Term Load Forecasting Based on a Semi-Parametric Additive Model. IEEE Transactions on Power Systems, 27(1), 134-141.\Sexpr[results=rd]{tools:::Rd_expr_doi("10.1109/TPWRS.2011.2162082")}.

See Also

model_smimodel, model_gaim, model_ppr, model_gam, model_lm

Examples

library(dplyr)
library(tibble)
library(tidyr)
library(tsibble)

# Simulate data
n = 1205
set.seed(123)
sim_data <- tibble(x_lag_000 = runif(n)) |>
  mutate(
    # Add x_lags
    x_lag = lag_matrix(x_lag_000, 5)) |>
  unpack(x_lag, names_sep = "_") |>
  mutate(
    # Response variable
    y = (0.9*x_lag_000 + 0.6*x_lag_001 + 0.45*x_lag_003)^3 + rnorm(n, sd = 0.1),
    # Add an index to the data set
    inddd = seq(1, n)) |>
  drop_na() |>
  select(inddd, y, starts_with("x_lag")) |>
  # Make the data set a `tsibble`
  as_tsibble(index = inddd)

# Training set
sim_train <- sim_data[1:1000, ]
# Validation set
sim_val <- sim_data[1001:1200, ]

# Predictors taken as non-linear variables
s.vars <- colnames(sim_data)[3:8]

# Model fitting
backwardModel <- model_backward(data = sim_train,
                                val.data = sim_val,
                                yvar = "y",
                                s.vars = s.vars)
# Fitted model
backwardModel$fit[[1]]

smimodel documentation built on April 8, 2026, 5:06 p.m.