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-HiTS extends N-BEATS with multi-rate sampling: different stacks operate at different temporal scales.

Each stack uses MLP blocks with downsampling/interpolation basis functions.

Stacks are summed hierarchically — coarse-to-fine — to produce the multi-horizon forecast.

Pros and cons on this universe

Pros

  • 50× cheaper than equivalent Transformers in the original paper.
  • Mid-tier 918-paper rank.
  • Confirmation arm to N-BEATS in our zoo.

Cons / failure modes

  • Rank #17 IC — beaten by BITCN, TiDE, and most of the mixer cluster.
  • RMSE-optimised; directional Sharpe is a separate question.
  • Cluster-redundant with N-BEATS.

References