Introducing DeepLOB: Deep Convolutional Neural Networks for Limit Order Books for after-work quants

Takeaway. DeepLOB treats a stack of limit-order-book snapshots as an “image”: convolutional filters (inception-style) scan local price–size patterns, then an LSTM reads the time sequence—microstructure as computer vision + sequence modelling.

DeepLOB: Deep Convolutional Neural Networks for Limit Order Books (Zihao Zhang, Stefan Zohren, Stephen Roberts, 2018; arXiv:1808.03668) is the after-work CNN-on-the-book classic. Pair with Sirignano’s spatial net.

Fit for learners: Must-read CNN+LSTM on LOB snapshots; perfect for learners who want to move from OHLCV to microstructure features in Python.

This Hub note is a plain-language intro. Not investment advice.

The problem

Hand-crafted book features (imbalance, queue sizes) miss local spatial patterns. Flattened MLPs ignore that level 1 sits next to level 2. Short-horizon labels are also brutally noisy once you include queue priority, latency, and fees.

The idea / core intuition

• Arrange T snapshots × (prices & sizes at many levels) into a 2D/3D tensor a CNN can scan.
• Inception-style filters capture multi-scale local structure; LSTM aggregates through time.
• Predict short-horizon mid-price direction (classification), with careful horizon definitions in the paper.
• Demand leakage-free event alignment; a shuffled tensor is cheating.

Research template—not a colocated strategy.

Why after-work Python learners care

This is the cleanest story for “why learn arrays before Keras”: if you cannot shape (T, 4*L) book features, you cannot implement DeepLOB. It is also a reason to respect time order.

Tiny evening exercise: build a tensor of shape (20, 40) of random book features (20 times × 10 levels × 4 fields); print x[0, :4] as “t=0, best-level fields”; write why permuting the time axis would leak. Learning only.

One-evening reading checklist

1. Sketch how a book snapshot becomes a CNN input.
2. Say what the LSTM is for after the convolutional stack.
3. List two reasons published accuracy ≠ tradable edge.
4. Contrast DeepLOB with Sirignano’s earlier spatial net.

Download PDF

Download PDF

Landing: arXiv abs

Important disclaimer

Not investment advice. This Hub note is educational only. DeepLOB is academic ML research—not a recommendation to trade limit-order-book neural nets with real capital.

What this is not

• Not CEF / subsidy framing
• Not Quant or Macro course CTA
• Not an HFT-as-a-service pitch

Next step

Single course CTA: SoloQuant Academy’s Introduction to Python Programming—indexing and nested data first; CNN tensors are just stricter arrays.