Poster Session: Deep Learning for VWAP Execution

This research presents a comprehensive framework for optimizing Volume Weighted Average Price (VWAP) execution in cryptocurrency markets using deep learning approaches. Through three interconnected studies, we demonstrate how moving beyond traditional volume curve prediction can enhance VWAP execution performance. First, we show how deep learning's automatic differentiation capabilities can directly optimize VWAP execution by minimizing either absolute or quadratic deviations from market VWAP. Our initial static model, implemented using a Temporal Linear Network architecture, consistently outperforms traditional volume-prediction approaches across multiple cryptocurrencies. Building on these results, we develop a dynamic execution framework utilizing recurrent neural networks that can adapt to changing market conditions during order execution. This dynamic approach further improves performance by incorporating real-time market feedback into the execution strategy. Finally, we introduce a novel multi-asset learning approach using signature-enhanced transformers, which enables efficient model deployment across multiple assets while maintaining superior performance. This architecture, which combines attention mechanisms with path signatures, demonstrates consistent improvement over both asset-specific and globally-fitted models for both previously seen and unseen assets. Our research not only advances the theoretical understanding of VWAP execution but also provides practical implementations that significantly improve trading efficiency in cryptocurrency markets.