N-BEATS and N-HiTS
Deep learning forecasting models built entirely from stacked, fully connected blocks rather than recurrent or convolutional layers, that decompose a time series into interpretable trend and seasonality pieces.
Many deep forecasting models rely on recurrent or attention layers built for sequence data, but N-BEATS showed that a much simpler architecture — a deep stack of ordinary fully connected layers, with no recurrence at all — can match or beat them on standard forecasting benchmarks. Each block in the stack predicts a piece of the series (say, a trend component or a seasonal component) and passes along a "residual" for the next block to explain, so the final forecast is the sum of interpretable pieces rather than one opaque output.
N-BEATS forecasts by stacking simple fully connected blocks that each explain one component of the series (trend, seasonality) and hand the leftover residual to the next block, producing both a strong forecast and an interpretable decomposition — no recurrence or attention required.
What N-HiTS adds
N-HiTS builds on the same block-stacking idea but adds multi-rate sampling: different blocks specialize in different time resolutions, with some blocks focused on long-run, low-frequency patterns and others on short-run, high-frequency detail. This hierarchical interpolation makes N-HiTS considerably faster and more accurate on long-horizon forecasts than the original N-BEATS, particularly for series many steps into the future.
Both models are popular in forecasting competitions precisely because they avoid recurrence entirely, making them fast to train while still offering some interpretability into what patterns the model is actually picking up.
Practice in interviews
Further reading
- Oreshkin, B. et al., 'N-BEATS: Neural Basis Expansion Analysis for Time Series Forecasting' (2020)
- Challu, C. et al., 'N-HiTS: Neural Hierarchical Interpolation for Time Series Forecasting' (2023)