N-HiTS Forecasting

N-HiTS (Neural Hierarchical Interpolation for Time Series) is a deep learning forecasting model that stacks multilayer perceptron blocks with multi-rate input pooling and hierarchical interpolation to forecast long horizons.

N-HiTS는 다중 비율 입력 풀링과 계층적 보간을 적용한 다층퍼셉트론 블록을 쌓아 긴 예측구간을 예측하는 딥러닝 시계열 예측모형이다.

How do I forecast a time series over a long horizon with deep learning?

딥러닝으로 긴 기간에 걸친 시계열을 어떻게 예측하나요?

N-HiTS forecasts time series with stacked neural network blocks that each model a different frequency band, combining their outputs through hierarchical interpolation. It fits long-horizon forecasting with substantial history, often trained across many related series.

When to use N-HiTS Forecasting

• Long forecast horizons with plenty of historical data
• Many related series that can be trained as one global model
• Patterns at several time scales

When to use something else

• Short series or few observations; use ARIMA, exponential smoothing or Prophet
• Interpretable components or effects of specific regressors are required; use SARIMAX or BSTS

Assumptions and requirements

• Enough history relative to the forecast horizon
• Scaling or normalization of series
• Time-ordered train/validation splits
• Regularly spaced observations

What the analysis reports

• Point forecasts and, with probabilistic losses, prediction intervals
• MAE, RMSE or MAPE on a held-out period
• Comparison against baseline forecasts
• Forecast plots

How to report it

N-HiTS achieved a test MAE of [value] over a [value]-step horizon, compared with [value] for a seasonal naive baseline.

Related methods

• N-BEATS
• Prophet Forecasting Model
• SARIMAX Model

Run N-HiTS Forecasting on Quorant