DeepLOB and CNN-LSTM Microstructure Models

A limit order book, snapshotted many times a second, is a grid of prices and volumes at each level on the bid and ask side, changing shape as orders arrive, trade, and cancel. Rather than hand-engineering features from that grid (imbalance ratios, spread, depth at each level), DeepLOB feeds the raw snapshots directly into a neural network and lets it learn which patterns in the shape matter for predicting short-horizon price moves.

The architecture pairs two components suited to two different jobs. A stack of convolutional layers scans across nearby price levels within a single snapshot, picking up local shape patterns — like a sudden buildup of volume just above the best ask — the same way a convolutional network scans across nearby pixels in an image. An LSTM layer then reads the sequence of these per-snapshot summaries across time, tracking how the book's shape evolves over the past several seconds, since order flow one second ago is informative about what happens next in a way a single static snapshot can't capture alone.

The combination — convolutions for spatial structure within one snapshot, an LSTM for temporal structure across snapshots — became a standard reference architecture for short-horizon (seconds-ahead) price-direction prediction directly from order book data, without needing a human to first design summary features by hand.