Forecast a time series with TimesFM

timesfm() produces a zero-shot probabilistic forecast of a univariate time series using the pretrained TimesFM 2.5 (200M) foundation model. There is no training step: the model reads your series as context and forecasts in a single forward pass.

Loading the pretrained weights is the slow part, so it happens once per session and the model is reused by later calls.

Usage

timesfm(y, horizon)

Arguments

y

A numeric vector or ts object: the observed series, oldest first, with no missing values (impute or drop them first). Series longer than 1024 points are used from the most recent 1024.

horizon

Number of future steps to forecast, between 1 and 256.

Value

An object of class "timesfm": a list with elements

• mean — numeric vector of point forecasts, length horizon.

• quantiles — numeric matrix with horizon rows and columns q10 ... q90: the deciles of the forecast distribution.

• horizon — the forecast length.

Examples

if (FALSE) { # \dontrun{
fc <- timesfm(AirPassengers, horizon = 24)
fc

fc$mean
fc$quantiles[, c("q10", "q90")]  # an 80% interval
} # }