Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

VWAP Execution with Signature-Enhanced Transformers: A Multi-Asset Learning Approach

Remi Genet

TL;DR

This work tackles VWAP execution challenges by introducing a globally-fitted transformer framework augmented with learnable path signatures to capture long-range, geometry-rich temporal patterns. The model, built on Temporal Kolmogorov-Arnold Transforms with a Variable Selection Network, demonstrates strong cross-asset generalization across 80 cryptocurrency pairs and outperforms asset-specific baselines, especially when signatures are included (GFT-Sig). Empirical results show significant reductions in both absolute and quadratic VWAP losses, with notable robustness to unseen assets and longer execution horizons. A real-time simulation confirms practical viability, indicating substantial improvements over TWAP and potential operational advantages for multi-asset trading desks.

Abstract

In this paper I propose a novel approach to Volume Weighted Average Price (VWAP) execution that addresses two key practical challenges: the need for asset-specific model training and the capture of complex temporal dependencies. Building upon my recent work in dynamic VWAP execution arXiv:2502.18177, I demonstrate that a single neural network trained across multiple assets can achieve performance comparable to or better than traditional asset-specific models. The proposed architecture combines a transformer-based design inspired by arXiv:2406.02486 with path signatures for capturing geometric features of price-volume trajectories, as in arXiv:2406.17890. The empirical analysis, conducted on hourly cryptocurrency trading data from 80 trading pairs, shows that the globally-fitted model with signature features (GFT-Sig) achieves superior performance in both absolute and quadratic VWAP loss metrics compared to asset-specific approaches. Notably, these improvements persist for out-of-sample assets, demonstrating the model's ability to generalize across different market conditions. The results suggest that combining global parameter sharing with signature-based feature extraction provides a scalable and robust approach to VWAP execution, offering significant practical advantages over traditional asset-specific implementations.

VWAP Execution with Signature-Enhanced Transformers: A Multi-Asset Learning Approach

TL;DR

This work tackles VWAP execution challenges by introducing a globally-fitted transformer framework augmented with learnable path signatures to capture long-range, geometry-rich temporal patterns. The model, built on Temporal Kolmogorov-Arnold Transforms with a Variable Selection Network, demonstrates strong cross-asset generalization across 80 cryptocurrency pairs and outperforms asset-specific baselines, especially when signatures are included (GFT-Sig). Empirical results show significant reductions in both absolute and quadratic VWAP losses, with notable robustness to unseen assets and longer execution horizons. A real-time simulation confirms practical viability, indicating substantial improvements over TWAP and potential operational advantages for multi-asset trading desks.

Abstract

In this paper I propose a novel approach to Volume Weighted Average Price (VWAP) execution that addresses two key practical challenges: the need for asset-specific model training and the capture of complex temporal dependencies. Building upon my recent work in dynamic VWAP execution arXiv:2502.18177, I demonstrate that a single neural network trained across multiple assets can achieve performance comparable to or better than traditional asset-specific models. The proposed architecture combines a transformer-based design inspired by arXiv:2406.02486 with path signatures for capturing geometric features of price-volume trajectories, as in arXiv:2406.17890. The empirical analysis, conducted on hourly cryptocurrency trading data from 80 trading pairs, shows that the globally-fitted model with signature features (GFT-Sig) achieves superior performance in both absolute and quadratic VWAP loss metrics compared to asset-specific approaches. Notably, these improvements persist for out-of-sample assets, demonstrating the model's ability to generalize across different market conditions. The results suggest that combining global parameter sharing with signature-based feature extraction provides a scalable and robust approach to VWAP execution, offering significant practical advantages over traditional asset-specific implementations.

Paper Structure

This paper contains 46 sections, 51 equations, 3 figures, 4 tables.

Table of Contents

  1. Introduction
  2. VWAP Definition and Discretization
  3. Classical VWAP Approaches
  4. Dynamic Volume Approaches
  5. The Rise of Deep Learning in Financial Time Series
  6. Deep Learning Approaches to Market Execution
  7. Path Signatures
  8. The Promise and Limitations of Attention Mechanisms
  9. Proposed Innovations
  10. Signature-Based Dynamic VWAP Architecture
  11. Model Overview
  12. Transformer-Based Architecture
  13. Volume Prediction
  14. Variable Selection Networks
  15. Initial Feature Embedding
  16. ...and 31 more sections

Figures (3)

  • Figure 1: Signature-Based Dynamic VWAP Architecture
  • Figure 2: Variable Selection Network Architecture
  • Figure 3: Gated Residual Network Architecture