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

RevIN normalises each lookback window: subtract the last value, divide by per-window std.

A single linear layer maps the normalised lookback to the normalised forecast.

RevIN is reversed at output to produce the un-normalised forecast.

Pros and cons on this universe

Pros

  • Cheapest possible competitive baseline — one linear layer.
  • Useful sanity check — if a deep model cannot beat NLinear-RevIN, the deep model is not contributing.
  • Reproducible and fast.

Cons / failure modes

  • Negative IC (−0.044) on this universe — below random.
  • Confirms the “floor” role but is not deployable as alpha.
  • Beaten by every structured architecture above rank 24 on the screen.

References