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

TiDE is a deep MLP with residual connections — no attention layers.

Lookback is compressed by an MLP encoder into a latent representation; an MLP decoder produces multi-horizon outputs from that latent.

Each layer is followed by a feature-wise linear transformation and a residual skip connection.

Pros and cons on this universe

Pros

  • Strongest BNB long signal in the audited table (+0.253 per-coin IC) and strong ETH (+0.198).
  • Pure MLP — cheap, parallelisable, hyperparameter-stable.
  • Confirms that attention is not necessary for competitive performance on this universe.

Cons / failure modes

  • Mid-rank #10 IC overall — beaten by the mixer family.
  • ~ρ ≈ +0.85 with BITCN on IC vectors — partial redundancy with the CNN family.

References