NPTS

Non-Parametric Time Series predictor predicts future values by sampling from past observations. The sampling weights can follow either a uniform or exponentially decreasing distribution, and optionally take into account the seasonality of the time series.

For a non-seasonal NPTS model, past observations can be sampled using a uniform distribution, or an exponential distribution whose probability decreases with the time distance to the observation (recent observations thus have a higher probability to be sampled than distant past ones).

For a seasonal NPTS model, time features based on the frequency of the time series are computed. Past observations can either be sampled using a uniform distribution amongst observations with the same values for the computed time features, or using an exponential distribution whose probability decreases with both the time distance to the past observation and the distance w.r.t computed time features. External features can also be used along with, or instead of, the computed time features.

Use default context length

Set the context length to the model's default value (context length = 1100).

Kernel type

Try Exponential Try Uniform

Sampling distribution of past observations.

Use a seasonal model

Use computed time features

Compute time features based on the frequency of the time series (minute of hour for minutely, hour of day for hourly...).

For a seasonal NPTS model, the computed time features are always used when no external features have been provided.

Random state

Using a fixed random seed allows for reproducible result.