CNN-LSTM on Binance order-book snapshots — experimental prototype

Status: learning-stage prototype, being rebuilt. The next iteration is a faithful DeepLOB (Zhang et al., 2019) replication validated on the FI-2010 benchmark, with chronological embargoed splits and cost-aware replay evaluation. Until that lands, everything here should be read as a prototype with the limitations listed below.

A two-process experiment: a C++ client polls Binance's public REST depth endpoint, normalizes the top 10 levels per side into 40 features, and ships them over ZeroMQ (raw float32 payload) to a Python server running a small CNN-LSTM classifier (DeepLOB_Simplified — a reduced stand-in, not the full DeepLOB architecture) that paper-trades the resulting signal.

Honest scope

• Spot REST polling, not WebSockets. Decision cadence is hundreds of milliseconds to seconds. This is not HFT, and no latency is measured anywhere in this repo.
• Paper trading only. There is no order execution. Fills are simulated at ticker price with zero fees, spread, or slippage, so any simulated PnL is optimistic by construction and no performance figures are published here.
• Known training-pipeline flaw. The committed preprocessing class-balances and shuffles rows before windowing and the train/test split, which scrambles the sequences and leaks across the split — reported validation accuracy is not meaningful. The rebuild replaces this with contiguous windows and chronological, embargoed splits.
• The solid piece: data_ingest.py — Binance depth20@100ms WebSocket collection to CSV with reconnect — carries over to the rebuild unchanged.

Pipeline

1. data_ingest.py — collect 100ms depth snapshots to CSV
2. preprocess_balanced.py — features + fixed-horizon labels
3. train.py — train DeepLOB_Simplified (saves model_balanced.pth)
4. strategy_server.py (Python, ZMQ REP) + bot.cpp (C++ client) — live signal loop, paper only

Build & run

# C++ client
cmake -B build && cmake --build build

# Python side
pip install -r requirements.txt
python strategy_server.py   # terminal 1
./build/hft_bot             # terminal 2