skaters for R

Port of skaters: online, univariate, distributional forecasting by conjugation. Pure R, no dependencies (jsonlite only for the parity tests).

Status: the full surface is ported, including laplace (the deployed default with GPD tails and the parade wrapper), the adaptive search, the spec builders, the periodicity detector, and the covariance estimators. Everything is parity-verified against the Python reference at 1e-6 (105,798 probe values), and two further gates run under R CMD check: the adversarial robustness streams and a determinism plus exact checkpoint-resume contract. See PORTING.md for the module map.

Installation

You can install the development version of skaters from GitHub with:

# install.packages("pak")
pak::pak("microprediction/skaters-r")

Example

library(skaters)

sk <- conjugate(leaf(1L), ema_transform(0.1), 1L)
state <- NULL
for (y in rnorm(100)) {
  r <- sk(y, state)
  state <- r$state
}
d <- r$dists[[1]]
c(dist_mean(d), dist_quantile(d, 0.9))