nnetar_reg() is a way to generate a specification of an NNETAR model before fitting and allows the model to be created using different packages. Currently the only package is forecast.

nnetar_reg(
  mode = "regression",
  seasonal_period = NULL,
  non_seasonal_ar = NULL,
  seasonal_ar = NULL,
  hidden_units = NULL,
  num_networks = NULL,
  penalty = NULL,
  epochs = NULL
)

Arguments

mode

A single character string for the type of model. The only possible value for this model is "regression".

seasonal_period

A seasonal frequency. Uses "auto" by default. A character phrase of "auto" or time-based phrase of "2 weeks" can be used if a date or date-time variable is provided. See Fit Details below.

non_seasonal_ar

The order of the non-seasonal auto-regressive (AR) terms. Often denoted "p" in pdq-notation.

seasonal_ar

The order of the seasonal auto-regressive (SAR) terms. Often denoted "P" in PDQ-notation.

hidden_units

An integer for the number of units in the hidden model.

num_networks

Number of networks to fit with different random starting weights. These are then averaged when producing forecasts.

penalty

A non-negative numeric value for the amount of weight decay.

epochs

An integer for the number of training iterations.

Details

The data given to the function are not saved and are only used to determine the mode of the model. For nnetar_reg(), the mode will always be "regression".

The model can be created using the fit() function using the following engines:

Main Arguments

The main arguments (tuning parameters) for the model are the parameters in nnetar_reg() function. These arguments are converted to their specific names at the time that the model is fit.

Other options and argument can be set using set_engine() (See Engine Details below).

If parameters need to be modified, update() can be used in lieu of recreating the object from scratch.

Engine Details

The standardized parameter names in modeltime can be mapped to their original names in each engine:

modeltimeforecast::nnetar
seasonal_periodts(frequency)
non_seasonal_arp (1)
seasonal_arP (1)
hidden_unitssize (10)
num_networksrepeats (20)
epochsmaxit (100)
penaltydecay (0)

Other options can be set using set_engine().

nnetar

The engine uses forecast::nnetar().

Function Parameters:

## function (y, p, P = 1, size, repeats = 20, xreg = NULL, lambda = NULL, 
##     model = NULL, subset = NULL, scale.inputs = TRUE, x = y, ...)

Parameter Notes:

  • xreg - This is supplied via the parsnip / modeltime fit() interface (so don't provide this manually). See Fit Details (below).

  • size - Is set to 10 by default. This differs from the forecast implementation

  • p and P - Are set to 1 by default.

  • maxit and decay are nnet::nnet parameters that are exposed in the nnetar_reg() interface. These are key tuning parameters.

Fit Details

Date and Date-Time Variable

It's a requirement to have a date or date-time variable as a predictor. The fit() interface accepts date and date-time features and handles them internally.

Seasonal Period Specification

The period can be non-seasonal (seasonal_period = 1 or "none") or yearly seasonal (e.g. For monthly time stamps, seasonal_period = 12, seasonal_period = "12 months", or seasonal_period = "yearly"). There are 3 ways to specify:

  1. seasonal_period = "auto": A seasonal period is selected based on the periodicity of the data (e.g. 12 if monthly)

  2. seasonal_period = 12: A numeric frequency. For example, 12 is common for monthly data

  3. seasonal_period = "1 year": A time-based phrase. For example, "1 year" would convert to 12 for monthly data.

Univariate (No xregs, Exogenous Regressors):

For univariate analysis, you must include a date or date-time feature. Simply use:

Multivariate (xregs, Exogenous Regressors)

The xreg parameter is populated using the fit() or fit_xy() function:

  • Only factor, ordered factor, and numeric data will be used as xregs.

  • Date and Date-time variables are not used as xregs

  • character data should be converted to factor.

Xreg Example: Suppose you have 3 features:

  1. y (target)

  2. date (time stamp),

  3. month.lbl (labeled month as a ordered factor).

The month.lbl is an exogenous regressor that can be passed to the nnetar_reg() using fit():

Note that date or date-time class values are excluded from xreg.

See also

Examples

library(dplyr) library(parsnip) library(rsample) library(timetk) library(modeltime) # Data m750 <- m4_monthly %>% filter(id == "M750") m750
#> # A tibble: 306 × 3 #> id date value #> <fct> <date> <dbl> #> 1 M750 1990-01-01 6370 #> 2 M750 1990-02-01 6430 #> 3 M750 1990-03-01 6520 #> 4 M750 1990-04-01 6580 #> 5 M750 1990-05-01 6620 #> 6 M750 1990-06-01 6690 #> 7 M750 1990-07-01 6000 #> 8 M750 1990-08-01 5450 #> 9 M750 1990-09-01 6480 #> 10 M750 1990-10-01 6820 #> # … with 296 more rows
# Split Data 80/20 splits <- initial_time_split(m750, prop = 0.8) # ---- NNETAR ---- # Model Spec model_spec <- nnetar_reg() %>% set_engine("nnetar") # Fit Spec set.seed(123) model_fit <- model_spec %>% fit(log(value) ~ date, data = training(splits))
#> frequency = 12 observations per 1 year
model_fit
#> parsnip model object #> #> Fit time: 251ms #> Series: outcome #> Model: NNAR(1,1,10)[12] #> Call: forecast::nnetar(y = outcome, p = p, P = P, size = size, repeats = repeats, #> xreg = xreg_matrix, decay = decay, maxit = maxit) #> #> Average of 20 networks, each of which is #> a 2-10-1 network with 41 weights #> options were - linear output units #> #> sigma^2 estimated as 0.0005868