Reproduce “Convolutional Attention in Betting Exchange Markets” on OpenResearch

About this paper

Rui Gonçalves, Vitor Miguel Ribeiro, Roman Chertovskih, António Pedro Aguiar

Researchers from the University of Porto developed an end-to-end automated trading system for betting exchange markets that integrates deep learning with novel convolutional attention and a roll padding method. This system achieved a 30.92% classification accuracy and generated a positive cumulative profit and loss in a 30-day simulated trading environment.

Key insights

• LSTM-based models, particularly when augmented with multi-head convolutional attention, are effective for capturing sequential dependencies in market depth data, outperforming other deep learning architectures.
• The proposed "roll padding" method and convolutional attention mechanisms positively impact the performance of deep learning models in multivariate time series classification by addressing data integrity and feature extraction challenges.
• Achieving a positive Profit & Loss in automated trading requires a holistic, end-to-end system that considers real-world market dynamics, such as market absorption capacity for stake sizing, beyond just classification accuracy.

Summary by alphaXiv.