lambdaTS {lambdaTS}

lambdaTS: Variational Seq2Seq Lambda Transformer Model for Time Series Analysis

Description

Time series analysis based on Lambda Transformer and Variational Seq2Seq, built on 'Torch'.

Usage

lambdaTS(
  data,
  target,
  future,
  past = future,
  ci = 0.8,
  deriv = 1,
  yjt = TRUE,
  shift = 0,
  smoother = FALSE,
  k_embed = 30,
  r_proj = ceiling(k_embed/3) + 1,
  n_heads = 1,
  n_bases = 1,
  activ = "linear",
  loss_metric = "elbo",
  optim = "adam",
  epochs = 30,
  lr = 0.01,
  patience = epochs,
  verbose = TRUE,
  sample_n = 100,
  seed = 42,
  dev = "cpu",
  starting_date = NULL,
  dbreak = NULL,
  days_off = NULL,
  min_set = future,
  holdout = 0.5,
  batch_size = 30
)

Arguments

data

A data frame with ts on columns and possibly a date column (not mandatory)

target

String. Time series names to be jointly analyzed within the seq2seq model

future

Positive integer. The future dimension with number of time-steps to be predicted

past

Positive integer. The past dimension with number of time-steps in the past used for the prediction. Default: future

ci

Confidence interval. Default: 0.8

deriv

Positive integer. Number of differentiation operations to perform on the original series. 0 = no change; 1: one diff; 2: two diff, and so on.

yjt

Logical. Performing Yeo-Johnson Transformation on data is always advisable, especially when dealing with different ts at different scales. Default: TRUE

shift

Vector of positive integers. Allow for target variables to shift ahead of time. Zero means no shift. Length must be equal to the number of targets. Default: 0.

smoother

Logical. Perform optimal smooting using standard loess. Default: FALSE

k_embed

Positive integer. Number of Time2Vec embedding dimensions. Minimum value is 2. Default: 30

r_proj

Positive integer. Number of dimensions for the reduction space (to reduce quadratic complexity). Must be largely less than k_embed size. Default: ceiling(k_embed/3) + 1

n_heads

Positive integer. Number of heads for the attention mechanism. Computationally expensive, use with care. Default: 1

n_bases

Positive integer. Number of normal curves to build on each parameter. Computationally expensive, use with care. Default: 1

activ

String. The activation function for the linear transformation of the attention matrix into the future sequence. Implemented options are: "linear", "leaky_relu", "celu", "elu", "gelu", "selu", "softplus", "bent", "snake", "softmax", "softmin", "softsign", "sigmoid", "tanh", "tanhshrink", "swish", "hardtanh", "mish". Default: "linear".

loss_metric

String. Loss function for the variational model. Two options: "elbo" or "crps". Default: "crps".

optim

String. Optimization methods available are: "adadelta", "adagrad", "rmsprop", "rprop", "sgd", "asgd", "adam". Default: "adam".

epochs

Positive integer. Default: 30.

lr

Positive numeric. Learning rate. Default: 0.01.

patience

Positive integer. Waiting time (in epochs) before evaluating the overfit performance. Default: epochs.

verbose

Logical. Default: TRUE

sample_n

Positive integer. Number of samples from the variational model to evalute the mean forecast values. Computationally expensive, use with care. Default: 100.

seed

Random seed. Default: 42.

dev

String. Torch implementation of computational platform: "cpu" or "cuda" (gpu). Default: "cpu".

starting_date

Date. Initial date to assign temporal values to the series. Default: NULL (progressive numbers).

dbreak

String. Minimum time marker for x-axis, in liberal form: i.e., "3 months", "1 week", "20 days". Default: NULL.

days_off

String. Weekdays to exclude (i.e., c("saturday", "sunday")). Default: NULL.

min_set

Positive integer. Minimun number for validation set in case of automatic resize of past dimension. Default: future.

holdout

Positive numeric. Percentage of time series for holdout validation. Default: 0.5.

batch_size

Positive integer. Default: 30.

Value

This function returns a list including:

• prediction: a table with quantile predictions, mean and std for each ts

• history: plot of loss during the training process for the joint-transformed ts

• plot: graph with history and prediction for each ts

• learning_error: errors for the joint-ts in the transformed scale (rmse, mae, mdae, mpe, mape, smape, rrse, rae)

• feature_errors: errors for each ts in the original scale (rmse, mae, mdae, mpe, mape, smape, rrse, rae)

• pred_stats: for each predicted time feature, IQR to range, KL-divergence, risk ratio, upside probability, averaged across time-points and compared at the terminal points.

• time_log

Author(s)

Giancarlo Vercellino [e-mail]

Examples

## Not run: 
lambdaTS(bitcoin_gold_oil, c("gold_close", "oil_Close"), 30, deriv = 1)

## End(Not run)