N-BEATS Forecasting

N-BEATS is a deep learning architecture for univariate time series forecasting that stacks fully connected blocks with backward and forward residual links, optionally decomposing forecasts into trend and seasonality.

N-BEATS는 역방향·순방향 잔차 연결을 가진 완전연결 블록을 쌓아 단변량 시계열을 예측하는 딥러닝 구조로, 예측을 추세와 계절성으로 분해할 수도 있다.

Can a deep learning model forecast a single time series without manual feature engineering?

딥러닝 모형이 수작업 변수 설계 없이 하나의 시계열을 예측할 수 있나요?

N-BEATS forecasts a univariate series by passing a lookback window through stacks of fully connected blocks, each removing the part of the signal it explains and adding its own forecast. It fits long or many related series, and its interpretable configuration separates trend and seasonal components.

When to use N-BEATS Forecasting

• Forecasting a univariate series with a long history
• Many related series that can train one shared model
• Interpretable trend and seasonality decomposition of forecasts

When to use something else

• Short series with few observations: use ARIMA or exponential smoothing
• External covariates drive the outcome: use a model with exogenous inputs such as ARIMAX or a temporal fusion transformer

Assumptions and requirements

• Regularly spaced series
• Enough history to fill many lookback windows
• Scaled or normalized inputs
• Evaluation on a held-out period after the training window

What the analysis reports

• Point forecasts for the horizon
• Error metrics (MAE, RMSE, MAPE or MASE)
• Comparison with a naive or statistical baseline
• Trend and seasonality components (interpretable version)

How to report it

On the held-out [period], N-BEATS achieved a MASE of [value] compared with [value] for the seasonal naive baseline.

Related methods

• ARIMA Model
• Exponential Smoothing
• Temporal Fusion Transformer

Run N-BEATS Forecasting on Quorant