Mechanism diagram · Hybrid / Mixer
MLP MIXER / HYBRID Input series lookback × N coins RevIN Patching (optional) Embed → d-model Time-mixing MLP mix across time tokens Feature-mixing MLP mix across channels Forecast head Forecast

How it works

StemGNN learns an inter-series graph adjacency matrix during training — explicitly modelling cross-variate dependencies.

It combines graph convolutional operations (across the learned graph) with spectral temporal blocks (frequency-domain).

Outputs from each spectral-temporal block are aggregated for the final forecast.

Pros and cons on this universe

Pros

  • Explicit cross-variate modelling — fits cross-sectional trading thesis.
  • Strong BNB long (+0.307 per-coin IC, ties TFT for best).
  • Rank #14 IC overall.

Cons / failure modes

  • Pairwise IC ρ = +0.999 with BITCN — extreme redundancy.
  • No dossier in the research base.
  • Learned graph can overfit on small universes.

References