Objective

ns_sdn.typing.Objective = Objective

Alias of Literal['mse', 'huber'].

Prediction loss on the scaled residual \(e_t = \hat y_t - y_t\).

"mse" is the published objective, \(\frac{1}{r}\sum_t e_t^2\). "huber" is the canonical specification’s robust alternative for series with jumps,

\[\begin{split}\frac{1}{r}\sum_t \begin{cases} \tfrac{1}{2} e_t^2, & |e_t| \le \delta,\\ \delta\bigl(|e_t| - \tfrac{1}{2}\delta\bigr), & |e_t| > \delta, \end{cases}\end{split}\]

with threshold \(\delta\) (delta) on the scaled series, so under min-max scaling \(\delta = 1\) is half the range of the sample.