Mechanism diagram · MLP / Linear
MLP / LINEAR Input lookback L flat features Dense layer with activation Dense layer + residual Linear head Linear floors (DLinear / NLinear-RevIN) collapse to one dense layer.

How it works

N-BEATS uses stacks of fully-connected blocks. Each block has two outputs: a backcast (reconstruction of the lookback) and a forecast (forward prediction).

The backcast is subtracted from the input before passing to the next block — doubly residual stacking.

Block basis functions can be generic (data-driven) or interpretable (polynomial trend, Fourier seasonality).

Pros and cons on this universe

Pros

  • Interpretable variant decomposes output into trend and seasonality components.
  • Strong baseline from the M4 competition era.
  • Rank #19 IC — still positive.

Cons / failure modes

  • Beaten by N-HiTS on the 918-paper crypto benchmark.
  • RMSE-focused — directional rank correlation is a side metric.

References