Principle:Sktime Pytorch forecasting NBeats Model Instantiation

Knowledge Sources	N-BEATS pytorch-forecasting
Domains	Time_Series, Deep_Learning, Signal_Decomposition
Last Updated	2026-02-08 07:00 GMT

Overview

Technique for instantiating the Neural Basis Expansion Analysis for Time Series (N-BEATS) model with interpretable trend and seasonal decomposition stacks.

Description

N-BEATS is a pure deep learning architecture for univariate time series forecasting that uses backward and forward residual connections with specialized basis expansion blocks. The model decomposes the input signal through a stack of blocks, each producing a partial forecast and a backcast (reconstruction of the input). The interpretable variant uses two stack types: trend blocks with polynomial basis functions and seasonality blocks with Fourier basis functions. This decomposition provides built-in interpretability — the model automatically learns to separate trend from seasonal components. N-BEATS operates on univariate data without covariates, using only the target history as input.

Usage

Use N-BEATS for univariate time series forecasting when: (1) no exogenous covariates are available or needed, (2) interpretable trend/seasonal decomposition is desired, (3) fixed encoder and decoder lengths are acceptable. N-BEATS requires min_encoder_length == max_encoder_length and min_prediction_length == max_prediction_length.

Theoretical Basis

N-BEATS processes the lookback window through a stack of blocks with residual connections:

Block operation: ${\hat{x}}_l=g_l^b(h_l),{\hat{y}}_l=g_l^f(h_l)$

Where $h_l$ is the block's hidden representation, $g^b$ is the backcast basis function, and $g^f$ is the forecast basis function.

Residual learning: $x_{l+1}=x_l-{\hat{x}}_l$

Basis functions for interpretable stacks:

• Trend: Polynomial basis $g^f(h)=\sum _{k=0}^p{\theta }_k{t}^k$
• Seasonality: Fourier basis $g^f(h)=\sum _{k=1}^K[a_k\cos (2\pi kt/T)+b_k\sin (2\pi kt/T)]$

Final forecast: Sum of all block forecasts: $\hat{y}=\sum _{l=1}^L{\hat{y}}_l$

Related Pages

Implemented By

• Implementation:Sktime_Pytorch_forecasting_NBeats_From_Dataset

Uses Heuristic

• Heuristic:Sktime_Pytorch_forecasting_Encoder_Decoder_Length_Limits