Contents

TimeSeriesForecasterBatches

A sequence of forecaster batches on a time series shaped array.

Declaration

struct TimeSeriesForecasterBatches<Scalar> where Scalar : MLShapedArrayScalar

Overview

A time-series forecaster takes a series of samples and produces a prediction of the next samples. For example the sequence [1, 2, 3, 4] could predict [5, 6]. To train a forecaster, each training batch contains the input samples along with the annotations (ground truth predictions). For example a batch could have this:

features = [
    [1, 2, 3, 4],
    [2, 3, 4, 5],
    [3, 4, 5, 6],
]
annotations = [
    [5, 6],
    [6, 7],
    [7, 8],
]

The shape of the features in the sequence is [batchSize, inputWindowSize, featureSize] and the shape of the annotations is [batchSize, forecastWindowSize, annotationSize]. The batch sequence will return as many feature-annotation examples as fit in the input. For example, an input sequence size of 10 with an input sample count of 4 and a prediction sample count of 2 will produce 5 examples:

features: [1, 2, 3, 4], annotations: [5, 6]
features: [2, 3, 4, 5], annotations: [6, 7]
features: [3, 4, 5, 6], annotations: [7, 8]
features: [4, 5, 6, 7], annotations: [8, 9]
features: [5, 6, 7, 8], annotations: [9, 10]

Note that 9 and 10 are never used as features because there would be no annotations for those examples.

Topics

Creating a time series forecaster batch

Inspecting a time series forecaster batch

Default Implementations

See Also

Time-based components