The Trick That Makes Deep Learning Actually Work on Order Books

Introduction

Order book imbalance predicts short-term returns. This is not news. What is news: the way you feed that data into your model matters far more than whether you use a fancy neural network.

A simple LSTM trained on stationary order flow achieves 1.2% out-of-sample R² on 115 Nasdaq stocks. The same LSTM trained on raw order book states? Basically zero. The CNN-LSTM that papers love to propose? Does marginally better on raw data, but once you transform inputs properly, it adds almost nothing.

Here’s what actually drives performance in high-frequency return prediction, and why most papers miss the point by obsessing over architecture.