•
Justin Sirignano (2016) applies spatial neural nets to limit-order-book states—an early LOB deep-learning baseline. Plain-language intro. Not investment advice.
Takeaway. A limit order book is a spatial object (price levels × sizes). Sirignano’s early paper treats that grid with spatial neural nets—a historical baseline before CNN+LSTM DeepLOB, not a latency-arb recipe.
Deep Learning for Limit Order Books (Justin Sirignano, 2016) is the after-work prequel to later LOB deep-learning hits.
Fit for learners: Early spatial neural net on LOB — historical baseline before DeepLOB for microstructure learners.
This Hub note is a plain-language intro. Not investment advice.
Tick-level books are huge, noisy, and local: nearby price levels interact. Flattening the book into a random forest of hand features throws away geometry that a spatial net might use.
Research pattern—not a co-location trading system.
Before DeepLOB’s CNN+LSTM stack, this paper answers: “why not treat the book like a 1D image?” Perfect once you can slice NumPy arrays.
Tiny evening exercise: make a toy book as two arrays of 10 sizes (bids, asks); print asks[0]-bids[0] as a fake spread; write why using the next snapshot’s mid as a feature would leak—learning only.
Landing: arXiv abs
Not investment advice. This Hub note is educational only. The paper is academic ML research—not a recommendation to trade order-book neural nets with real capital.
Single course CTA: SoloQuant Academy’s Introduction to Python Programming—arrays and loops first; book tensors become a weekend project afterward.