Deep Learning On Limit Order Books

A limit order book is a live, high-dimensional snapshot: prices and quantities stacked at each level on the bid and ask sides, updating with every order, cancel, and trade. Classical microstructure models compress that snapshot into a handful of hand-built features — order imbalance, spread, queue depth at the best level — before feeding them to a simple model. Deep learning approaches instead feed the raw stacked-level data directly into a neural network and let it learn which combinations of levels and updates matter, rather than deciding that in advance.

The common architecture treats the order book as a 2D grid, price levels across one axis and time across the other, and applies convolutional layers to pick up local patterns — a thinning ask side just above the touch, say — the same way a convolutional network finds edges in an image. Recurrent or transformer layers on top capture how those patterns evolve over the last few seconds. The output is typically a short-horizon prediction: will the mid-price move up, down, or stay flat in the next few hundred milliseconds.

These models tend to outperform hand-built features in backtests but are notoriously fragile in production: a network trained on one stock's queue dynamics or one volatility regime often degrades sharply when the venue, tick size, or regime changes, and the "universal features" claim from early papers has held up only partially across markets.