LOB-specialised

DeepLOB

CNN + Inception + LSTM stack on raw L2 order-book snapshots; the canonical LOB classifier.

Status · frontier
WF IC (shared metric) · not yet measured
Status · Dossier only · gated on L2 backfill EXP-006
Mechanism diagram · LOB-specialised
LOB-SPECIALISED (DeepLOB-style) L2 order book T snapshots × 40 cols Conv 1×4 price/qty pairs Conv 1×2 ladder fusion Inception multi-scale LSTM (small) temporal pooling Softmax 3-way up/flat/down

How it works

DeepLOB takes a sequence of L2 order-book snapshots (40 columns × 100 timesteps in the original).

Layer 1: 1×4 convolutions across price/quantity pairs at each level.

Layer 2: 1×2 convolutions fuse adjacent levels.

Inception-style multi-scale convolutions follow.

A small LSTM pools the temporal dimension.

Final softmax classifies the next price move into up / flat / down.

Pros and cons on this universe

Pros

  • Famously reported ~70 percent directional accuracy on London Stock Exchange data.
  • Native handling of L2 ladder structure that candle-based architectures miss.
  • Strong literature footprint.

Cons / failure modes

  • No HL empirical IC — gated on the EXP-006 L2 backfill completing.
  • LOBCAST critique (Briola et al. 2023) shows poor generalisation off FI-2010.
  • Even at 70 percent directional accuracy, the underlying paper does not include realistic transaction costs — a known weakness in the LOB-classification literature.

References