metrics.make_mase()

Build a Mean Absolute Scaled Error metric scaled by train_data.

Usage

Source

metrics.make_mase(
    train_data,
    *,
    seasonality=1,
)

MASE divides the forecast MAE (using the sample median as point estimate) by the in-sample MAE of the seasonal-naive forecast on train_data, mean(|y_t - y_{t-seasonality}|). The scale is computed once at factory time; the returned metric has the standard scalar-array signature (see ~numpyro_forecast.typing.Metric).

Parameters

train_data: Float[Array, "*batch time obs_dim"]

Training data with time at axis -2; leading batch axes are allowed and the seasonal-naive scale is averaged over all axes.

seasonality: int = 1

Seasonal period (>= 1); 1 is the random-walk naive baseline.

Returns

Metric

A (pred, truth) callable computing MASE as a scalar array.

Raises

ValueError

If seasonality < 1, train_data is not longer than seasonality along the time axis, or the seasonal-naive scale is zero (a constant series, for which MASE is undefined).