N-HiTS
Neural Hierarchical Interpolation for Time Series Forecasting · Also known as: N-HiTS — Hiyerarşik İnterpolasyon Tahmini, NHITS, Neural Hierarchical Interpolation
N-HiTS (Neural Hierarchical Interpolation for Time Series Forecasting), introduced by Challu and colleagues in 2023, is a deep neural forecasting architecture that combines the hierarchical forecasts of multiple stacks operating at different sampling rates and merges them through interpolation. It extends N-BEATS to deliver markedly better accuracy on long forecast horizons.
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-HiTS for long-horizon forecasting and prediction on a continuous univariate time series when you have ample history (at least about 100 observations, and ideally several hundred). It is assumption-light regarding distribution but expects enough historical data, benefits from a GPU, and is most worthwhile when the forecast horizon is long. On very short series simpler classical models such as ARIMA are safer.
Strengths & limitations
- Strong accuracy on long forecast horizons, improving on N-BEATS.
- Multi-rate sampling lets the model capture both slow trends and fast fluctuations.
- Hierarchical interpolation keeps parameters and compute manageable even for long horizons.
- No normality assumption on the data.
- Needs ample historical data; on small time series (n below about 500) the hierarchical structure tends to overfit.
- On very short series (n below about 100) the neural hierarchical model is not meaningful.
- A GPU is recommended, raising the compute and setup cost relative to classical models.
- Like other deep models it offers little explicit interpretability of its parameters.
Frequently asked
How is N-HiTS different from N-BEATS?
N-HiTS builds on N-BEATS by adding multi-rate input sampling and hierarchical interpolation across stacks, which improves accuracy and efficiency specifically on long forecast horizons.
How much data do I need?
It expects ample history — at least about 100 observations as a floor, and ideally several hundred. Below roughly 500 points the hierarchical structure tends to overfit, and below about 100 the model is not meaningful.
Do I need a GPU?
A GPU is recommended. Like other deep forecasting models, N-HiTS trains far more comfortably with GPU acceleration, especially for long horizons.
What should I use on a short time series instead?
For short series, classical methods are safer: ARIMA when the series is very short, or a model such as random forest when the dataset is modest in size.
Sources
- Challu, C. et al. (2023). NHITS: Neural Hierarchical Interpolation for Time Series Forecasting. AAAI. DOI: 10.1609/aaai.v37i6.25854 ↗
- Oreshkin, B.N. et al. (2020). N-BEATS: Neural Basis Expansion Analysis for Interpretable Time Series Forecasting. ICLR. arXiv: 1905.10437 link ↗
How to cite this page
ScholarGate. (2026, June 1). Neural Hierarchical Interpolation for Time Series Forecasting. ScholarGate. https://scholargate.app/en/deep-learning/nhits
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.
Compare side by side →