The goal of timesfm is to make TimesFM, Google Research’s foundation model for time series, easy to use from R. You give it an observed series and a horizon; it gives you point forecasts and forecast deciles with no training, no model selection, and no tuning — the pretrained model reads your series as context and forecasts in a single forward pass.
It is a sibling of the tabfm package (zero-shot classification/regression on tabular data), built the same way: a thin reticulate wrapper around the official Python package.
You can install the development version of timesfm like so:
# install.packages("remotes")
remotes::install_github("mattyoreilly/timesfm")Then, one time only, install the underlying Python package into a dedicated virtualenv:
library(timesfm)
install_timesfm()Pass a numeric vector or ts, oldest observation first, and a horizon
of up to 256 steps:
library(timesfm)
fc <- timesfm(AirPassengers, horizon = 24)
fc
fc$mean # point forecasts
fc$quantiles[, c("q10", "q90")] # an 80% prediction intervalThe first call loads the pretrained weights (downloaded once from the Hugging Face Hub and cached); after that, forecasts in the same session are fast.
The pretrained weights are cached after the first download, but one
optional line in your ~/.Renviron (no Hugging Face account needed)
makes cold starts noticeably faster:
HF_HUB_OFFLINE=1
It loads weights straight from the local cache instead of revalidating them against the Hugging Face Hub on each session’s first forecast. Unset it whenever you need to download weights you don’t have yet.
- Forecasts require a working Python installation;
install_timesfm()sets everything up via reticulate. - Univariate series only, context capped at the most recent 1024 points, horizon capped at 256 steps.
- TimesFM is zero-shot: there is nothing to tune. If you need covariates, hierarchical structure, or bespoke accuracy on one series, reach for fable or prophet.