Deep Bidding Network: Learning Market Microstructure Dynamics from Limit Order Books

Abstract

Accurate modeling of stock market microstructure is foundational to algorithmic trading and market surveillance. Traditional approaches rely on handcrafted features extracted from limit order books, often failing to capture temporal dynamics that influence short-term price formation. In this work, a deep neural network architecture termed Deep Bidding Network (DBN) is introduced to learn latent representations of limit order book evolution. Unlike recurrent architectures that treat time series uniformly, DBN hierarchically encodes microsecond-level price changes and cross-asset interaction signals using convolutional flows. Evaluation on high-frequency trading data from the NASDAQ exchange over a three-month period showed that DBN achieves a 12.8% reduction in prediction error for mid-price movement compared to benchmark models. Risk metrics derived from the model’s output also produced sharper volatility estimates under stress market conditions. The study sheds light on the value of learned features in microstructure dynamics, though further work is necessary to assess robustness across diverse market regimes.