Mechanism diagram · CNN / TCN
CNN / TCN (DILATED) Input series lookback × N Conv 1×k dilation 1 Conv 1×k dilation 2 Conv 1×k dilation 4 Residual + LayerNorm stacked depth-wise blocks Pool + head Forecast

How it works

BITCN runs two TCN stacks in parallel — one causal (sees only past), one reverse-causal (sees only future within the lookback window).

The two streams are fused before the forecast head — bidirectional information without breaking the lookback boundary.

Dilated convolutions expand the receptive field exponentially with depth.

Pros and cons on this universe

Pros

  • Campaign-best BNB long (+0.298) and ETH long (+0.239) per-coin IC.
  • Not in the literature top-20 — a useful counter-example to literature priors.
  • Rank #11 IC.

Cons / failure modes

  • Pairwise IC ρ = +0.999 with StemGNN — almost perfectly redundant.
  • No dossier in the research base.
  • Bidirectional design slightly more expensive than causal TCN.

References