N-BEATS and N-HiTS

N-BEATS and N-HiTS are neural forecasting architectures built around direct multi-step forecasts from fixed history windows. They are not generic sequence encoders like RNN and LSTM forecasting or attention models. Their distinctive idea is to stack blocks that repeatedly explain part of the input history and contribute part of the forecast.

In N-BEATS, each block receives a residual backcast input $x$, passes it through fully connected layers, and emits two vectors: a backcast $\hat{x}$ that tries to explain the input window, and a forecast $\hat{y}$ for the horizon. The next block receives the residual $x-\hat{x}$, while forecasts from blocks are added:

$$\hat{y}=\sum _b{\hat{y}}^{(b)}.$$

Some N-BEATS stacks use generic learned bases. Interpretable stacks constrain the basis so one part behaves like trend and another like seasonality. The residual stacking is the mechanism: later blocks focus on structure earlier blocks did not explain.

N-HiTS keeps the residual-stack idea but changes how blocks handle resolution. It uses hierarchical interpolation and multi-rate sampling so different stacks specialize in different frequency bands. Coarser blocks can model slow long-horizon movement cheaply; finer blocks can model near-term detail. That design targets a practical weakness of plain fully connected horizon models: long horizons can be expensive and can mix low-frequency and high-frequency signals in one representation.

These models usually work best as global models across many related series with careful scaling, window construction, and horizon-specific evaluation. They are poor magic defaults for small single series where statistical forecasting, ARIMA, or exponential smoothing already fit the structure. If probabilistic outputs are needed, the loss and output head must be chosen accordingly, for example quantile outputs tied to quantile loss.

Connections

N-BEATS and N-HiTS are specialized deep learning forecasting models. They compete with temporal convolutional networks and transformer-based forecasting, and their value should be established through forecast error metrics by horizon and segment.

References

• Oreshkin et al., N-BEATS
• Challu et al., N-HiTS
• Nixtla NeuralForecast model documentation

Nav

Section — Time-Series Forecasting

← Transformer-Based Forecasting Forecast Evaluation →