Dynamic Portfolio Construction in High-Frequency Markets Using Microstructure-Aware Deep Temporal Models

Abstract

High-frequency financial markets exhibit rapidly changing liquidity and strong nonstationarity, which reduce the effectiveness of portfolio methods based on static return and covariance estimates. This study presents a microstructure-aware framework for dynamic portfolio allocation that maps high-frequency market states directly to portfolio weights. The model uses order-flow intensity, depth imbalance, and volume asymmetry as key inputs and combines recurrent and convolutional temporal components to capture both persistent dependencies and short-lived microstructure shocks. Portfolio weights are determined under a joint return–risk objective. The framework is evaluated on ten years of high-frequency U.S. equity data. Results show higher risk-adjusted performance than mean–variance and risk-parity benchmarks, with a 14%–22% improvement in Sharpe ratio and more stable drawdown behavior. These findings suggest that integrating market microstructure information into temporal allocation models can improve exposure management in intraday portfolio optimization.