DeepARScaleHandling

Enums · AiDotNet.Enums

Selects how DeepAR rescales a series before the recurrent trunk sees it.

For Beginners

Real series live on wildly different magnitudes - one stock trades at $10 and another at $1000, one product sells five units a day and another five thousand. A network that sees the raw numbers spends its capacity learning the magnitudes instead of the shapes. DeepAR divides each series by a single number that summarizes its own level, learns on the result, and multiplies the prediction back afterwards, so every series arrives at the network looking roughly the same size.

How It Works

Salinas et al. 2020, section 3.3 (“Scale handling”) defines that number as nu = 1 + mean(z) over the conditioning range, which is PaperMean and the default. The alternatives exist because that definition assumes non-negative data: a series that is already centered on zero, such as a return series, has a mean near zero and gains nothing from it.

Fields

Field Summary
DeepARScaleHandling MeanAbsolute nu = mean(|z|) over the conditioning range.
DeepARScaleHandling None No rescaling: the network sees the series as supplied.
DeepARScaleHandling PaperMean The paper’s scale factor, nu = 1 + mean(z) over the conditioning range (Salinas et al.
int value__