NLinearOptions<T>

Options & Configuration · AiDotNet.Models.Options

Options for NLinearModel — the normalization-linear forecaster (Zeng et al., AAAI 2023). It subtracts the last observed value from the window, applies a single linear map, then adds the value back — a distribution-shift-robust baseline that, like DLinear, is a strong modern control against heavier models.

Inheritance: ModelOptions → RegressionOptions<T> → TimeSeriesRegressionOptions<T> → NLinearOptions<T>

For Beginners

this model first “re-centers” the recent history around its most recent point, learns a straight-line mapping to the future, then shifts the prediction back up by that same recent value. The re-centering trick helps a lot when the overall level of the series drifts over time.

How It Works

NLinear is the simplest member of the LTSF-Linear family: normalize the input window by subtracting its last value, run one linear layer, then add the last value back to the output. That subtract/add step makes the model robust to distribution shift between training and test (a common failure mode of transformers on long-horizon forecasting), while keeping the model to a single linear map.

Typical sizes: a LookbackWindow of 24–96 steps with ForecastHorizon 1; the parameter count is about LookbackWindow × ForecastHorizon — a few thousand weights at most.

Reference: Ailing Zeng, Muxi Chen, Lei Zhang, Qiang Xu. “Are Transformers Effective for Time Series Forecasting?” AAAI 2023. https://arxiv.org/abs/2205.13504

Constructors

Constructor Summary
NLinearOptions() Creates a new NLinearOptions with default values.
NLinearOptions(NLinearOptions<T> other) Creates a deep copy of an existing NLinearOptions.

Properties

Property Summary
double LearningRate { get; set; } Step size for gradient-descent optimization.
int BatchSize { get; set; } Number of training samples processed per gradient update.
int Epochs { get; set; } Number of full passes over the training data.
int ForecastHorizon { get; set; } Forecast horizon.
int LookbackWindow { get; set; } History length the linear map sees (input window).