lambdaTS 2.0: probabilistic multivariate forecasting

Giancarlo Vercellino

Overview

lambdaTS fits a variational sequence-to-sequence model for jointly forecasting multiple numeric time series. Version 2.0 keeps the original lambdaTS() API, uses internal preprocessing helpers, and produces predictive samples and interval summaries.

Forecasting

library(lambdaTS)

result <- lambdaTS(
  data = bitcoin_gold_oil,
  target = c("gold_close", "oil_Close"),
  future = 10,
  past = 30,
  deriv = 1,
  epochs = 5,
  sample_n = 50,
  seed = 42
)

The returned prediction list contains horizon-by-horizon quantiles, means, standard deviations, minima, and maxima. feature_errors reports validation metrics on the original scale, while history and plot provide visual diagnostics.

Reproducibility

Set seed for reproducible preprocessing and torch initialization. The model can use dev = "cuda" when a compatible torch installation and GPU are available; CPU is the default.