N-BEATS
N-BEATS (Neural Basis Expansion Analysis for Interpretable Time Series Forecasting) · Also known as: N-BEATS — Nöral Zaman Serisi Tahmini, Neural Basis Expansion Analysis, neural basis expansion
N-BEATS is a deep learning architecture for time series forecasting, introduced by Oreshkin and colleagues in 2020, built from interpretable trend and seasonality stacks. It was the first purely neural forecasting model to reach state-of-the-art performance on the M4 competition without relying on any classical statistical components.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use N-BEATS for univariate, continuous time series forecasting when you have a reasonably long history (at least about 100 observations, and ideally several hundred) and want strong predictive accuracy. You must fix the forecast horizon and the input window in advance, and a GPU is recommended for training. On short series (below about 500 points) the stacked architecture tends to overfit, and below roughly 100 points a neural forecaster is meaningless — a classical statistical method such as ARIMA is the safer choice.
Strengths & limitations
- Reached state-of-the-art accuracy on the M4 competition as the first purely neural forecaster, with no statistical preprocessing.
- Interpretable trend-and-seasonality configuration makes the forecast decomposition readable.
- Distribution-light: does not require normally distributed data.
- Residual stacking gives the network capacity to capture complex temporal patterns.
- Needs a reasonably long history; on small series (below about 500 points) the stack overfits.
- On very short series (below about 100 points) a neural forecaster is meaningless and a classical method is preferable.
- Forecast horizon and input window must be chosen in advance.
- A GPU is recommended, adding compute and setup overhead.
Frequently asked
How much data does N-BEATS need?
It is designed for series with a reasonably long history — at least about 100 observations, and ideally several hundred. Below roughly 500 points the stacked blocks tend to overfit, and below about 100 points a neural forecaster is not meaningful.
Is N-BEATS interpretable?
In its interpretable configuration the blocks are grouped into a trend stack and a seasonality stack, so the final forecast is the sum of a trend part and a seasonality part. This decomposition is what makes the prediction readable rather than a black box.
When should I use ARIMA instead?
On very short series — below about 100 observations — a classical statistical method such as ARIMA is the safer choice, because a neural forecaster has too little signal to learn from and will overfit.
What do I need to set before training?
You must fix the forecast horizon and the input (look-back) window in advance, and a GPU is recommended for training within practical time.
Sources
- Oreshkin, B.N. et al. (2020). N-BEATS: Neural Basis Expansion Analysis for Interpretable Time Series Forecasting. ICLR. link ↗
- Makridakis, S., Spiliotis, E. & Assimakopoulos, V. (2020). The M4 Competition: 100,000 Time Series and 61 Forecasting Methods. International Journal of Forecasting, 36(1), 54–74. DOI: 10.1016/j.ijforecast.2019.04.014 ↗
How to cite this page
ScholarGate. (2026, June 1). N-BEATS (Neural Basis Expansion Analysis for Interpretable Time Series Forecasting). ScholarGate. https://scholargate.app/en/deep-learning/nbeats
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- ARIMAEconometrics↔ compare
- DeepARDeep learning↔ compare
- InformerDeep learning↔ compare
- Random ForestMachine learning↔ compare
- Temporal Fusion TransformerDeep learning↔ compare