nbeats: General Interface for N-BEATS Time Series Models
In modeltime.gluonts: 'GluonTS' Deep Learning

Description Usage Arguments Details Engine Details Engine Fit Details See Also Examples

nbeats() is a way to generate a specification of a N-BEATS model before fitting and allows the model to be created using different packages. Currently the only package is gluonts. There are 2 N-Beats implementations: (1) Standard N-Beats, and (2) Ensemble N-Beats.

nbeats(
  mode = "regression",
  id,
  freq,
  prediction_length,
  lookback_length = NULL,
  loss_function = NULL,
  bagging_size = NULL,
  num_stacks = NULL,
  num_blocks = NULL,
  epochs = NULL,
  batch_size = NULL,
  num_batches_per_epoch = NULL,
  learn_rate = NULL,
  learn_rate_decay_factor = NULL,
  learn_rate_min = NULL,
  patience = NULL,
  clip_gradient = NULL,
  penalty = NULL
)

`mode`	A single character string for the type of model. The only possible value for this model is "regression".
`id`	A quoted column name that tracks the GluonTS FieldName "item_id"
`freq`	A `pandas` timeseries frequency such as "5min" for 5-minutes or "D" for daily. Refer to Pandas Offset Aliases.
`prediction_length`	Numeric value indicating the length of the prediction horizon
`lookback_length`	Number of time units that condition the predictions Also known as 'lookback period'. Default is 2 * prediction_length.
`loss_function`	The loss function (also known as metric) to use for training the network. Unlike other models in GluonTS this network does not use a distribution. One of the following: "sMAPE", "MASE" or "MAPE". The default value is "MAPE".
`bagging_size`	(Applicable to Ensemble N-Beats). The number of models that share the parameter combination of 'context_length' and 'loss_function'. Each of these models gets a different initialization random initialization. Default and recommended value: 10.
`num_stacks`	The number of stacks the network should contain. Default and recommended value for generic mode: 30 Recommended value for interpretable mode: 2
`num_blocks`	The number of blocks per stack. A list of ints of length 1 or 'num_stacks'. Default and recommended value for generic mode: 1. Recommended value for interpretable mode: 3.
`epochs`	Number of epochs that the network will train (default: 5).
`batch_size`	Number of examples in each batch (default: 32).
`num_batches_per_epoch`	Number of batches at each epoch (default: 50).
`learn_rate`	Initial learning rate (default: 10-3).
`learn_rate_decay_factor`	Factor (between 0 and 1) by which to decrease the learning rate (default: 0.5).
`learn_rate_min`	Lower bound for the learning rate (default: 5x10-5 ).
`patience`	The patience to observe before reducing the learning rate, nonnegative integer (default: 10).
`clip_gradient`	Maximum value of gradient. The gradient is clipped if it is too large (default: 10).
`penalty`	The weight decay (or L2 regularization) coefficient. Modifies objective by adding a penalty for having large weights (default 10-8 ).

These arguments are converted to their specific names at the time that the model is fit. Other options and arguments can be set using set_engine(). If left to their defaults here (see above), the values are taken from the underlying model functions. If parameters need to be modified, update() can be used in lieu of recreating the object from scratch.

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

GluonTS N-BEATS: "gluonts_nbeats" (the default)
GluonTS N-BEATS Ensemble: "gluonts_nbeats_ensemble"

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

modeltime	NBEATSEstimator	NBEATSEnsembleEstimator
id	ListDataset('item_id')	ListDataset('item_id')
freq	freq	freq
prediction_length	prediction_length	prediction_length
lookback_length	context_length (= 2 x prediction_length)	meta_context_length (= prediction_length x c(2,4))
bagging_size	NA	meta_bagging_size (3)
loss_function	loss_function ('sMAPE')	meta_loss_function (list('sMAPE'))
num_stacks	num_stacks (30)	num_stacks (30)
num_blocks	num_blocks (list(1))	num_blocks (list(1))
epochs	epochs (5)	epochs (5)
batch_size	batch_size (32)	batch_size (32)
num_batches_per_epoch	num_batches_per_epoch (50)	num_batches_per_epoch (50)
learn_rate	learning_rate (0.001)	learning_rate (0.001)
learn_rate_decay_factor	learning_rate_decay_factor (0.5)	learning_rate_decay_factor (0.5)
learn_rate_min	minimum_learning_rate (5e-5)	minimum_learning_rate (5e-5)
patience	patience (10)	patience (10)
clip_gradient	clip_gradient (10)	clip_gradient (10)
penalty	weight_decay (1e-8)	weight_decay (1e-8)

Other options can be set using set_engine().

gluonts_nbeats

The engine uses gluonts.model.n_beats.NBEATSEstimator(). Default values that have been changed to prevent long-running computations:

epochs = 5: GluonTS uses 100 by default.
loss_function = 'sMAPE': GluonTS by default uses MAPE. MAPE can suffer from issues with small values.

Required Parameters

The gluonts_nbeats implementation has several Required Parameters, which are user-defined.

1. ID Variable (Required):

An important difference between other parsnip models is that each time series (even single time series) must be uniquely identified by an ID variable.

The ID feature must be of class character or factor.
This ID feature is provided as a quoted expression during the model specification process (e.g. nbeats(id = "ID") assuming you have a column in your data named "ID").

2. Frequency (Required):

The GluonTS models use a Pandas Timestamp Frequency freq to generate features internally. Examples:

freq = "5min" for timestamps that are 5-minutes apart
freq = "D" for Daily Timestamps

The Pandas Timestamps are quite flexible. Refer to Pandas Offset Aliases.

3. Prediction Length (Required):

Unlike other parsnip models, a prediction_length is required during the model specification and fitting process.

gluonts_nbeats_ensemble

The engine uses gluonts.model.n_beats.NBEATSEnsembleEstimator().

Number of Models Created

This model is very good, but can be expensive (long-running) due to the number of models that are being created. The number of models follows the formula:

length(lookback_length) x length(loss_function) x meta_bagging_size

The default values that have been changed from GluonTS implementation to prevent long-running computations:

epochs = 5: GluonTS uses 100 by default.
lookback_length = prediction_length * c(2, 4). GluonTS uses range of 2:7, which doubles the number of models created.
bagging_size = 3: Averages 5 like models together. GluonTS uses 10, which doubles the number of models created.
loss_function = 'sMAPE': GluonTS uses 3 meta_loss_function = list('sMAPE', 'MASE', 'MAPE'), which 3X's (triples) the number of models created.

The result is: 2 x 1 x 3 = 6 models. Each model will have 5 epochs by default.

Required Parameters

The gluonts_nbeats_ensemble implementation has several Required Parameters, which are user-defined.

1. ID Variable (Required):

An important difference between other parsnip models is that each time series (even single time series) must be uniquely identified by an ID variable.

The ID feature must be of class character or factor.
This ID feature is provided as a quoted expression during the model specification process (e.g. nbeats(id = "ID") assuming you have a column in your data named "ID").

2. Frequency (Required):

The GluonTS models use a Pandas Timestamp Frequency freq to generate features internally. Examples:

freq = "5min" for timestamps that are 5-minutes apart
freq = "D" for Daily Timestamps

The Pandas Timestamps are quite flexible. Refer to Pandas Offset Aliases.

3. Prediction Length (Required):

Unlike other parsnip models, a prediction_length is required during the model specification and fitting process.

The following features are REQUIRED to be available in the incoming data for the fitting process.

Fit: fit(y ~ date + id, data): Includes a target feature that is a function of a "date" and "id" feature. The ID feature must be pre-specified in the model_specification.
Predict: predict(model, new_data) where new_data contains both a column named "date" and "id".

ID Variable

An ID feature must be included in the recipe or formula fitting process. This assists with cataloging the time series inside GluonTS ListDataset. The column name must match the quoted feature name specified in the nbeats(id = "id") expects a column inside your data named "id".

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.

fit.model_spec(), set_engine()

library(tidymodels)
library(tidyverse)
library(timetk)


# ---- MODEL SPEC ----
# - Important: Make sure *required* parameters are provided
model_spec <- nbeats(

    # User Defined (Required) Parameters
    id                    = "id",
    freq                  = "M",
    prediction_length     = 24,

    # Hyper Parameters
    epochs                = 1,
    num_batches_per_epoch = 4
) %>%
    set_engine("gluonts_nbeats")

model_spec

# ---- TRAINING ----
# Important: Make sure the date and id features are included as regressors
#  and do NOT dummy the id feature.
model_fitted <- model_spec %>%
    fit(value ~ date + id, m750)

model_fitted

# ---- PREDICT ----
# - IMPORTANT: New Data must have id and date features
new_data <- tibble(
    id   = factor("M750"),
    date = as.Date("2015-07-01")
)

predict(model_fitted, new_data)