Application of Deep Learning in Multi-Agent Financial Simulation

Supervised by Professor Peng Yang, I wrote a paper as a co-author, SimLOB: Learning Representations of Limited Order Book for Financial Market Simulation, and submitted it to NeurIPS. You can find the preprint here:https://arxiv.org/abs/2406.19396.

We are the first to use LOB data to calibrate the PGPS model (previously, most studies relied on midprice). During the calibration process, to address the challenges of measuring similarity in time series data, we employed a representation learning approach. Specifically, we used a trained encoder to map the raw data into a latent space, where similarity metrics were computed for calibration. We also explored suitable encoder architectures, comparing contemporary common architectures and those frequently used in financial calibration, and designed a novel transformer-based architecture that significantly improved calibration performance.

What I did

• I built an autoencoder based on the transformer and designed a new calibration method to calibrate the multi-agent LOB financial simulator to address lob data representation and spatiotemporal calibration difficulties, which improved by more than 50%.
• I designed multi-angle experiments in the field to compare with existing algorithms like TransLOB and DepLOB, designed various visualizations to show the results, and constructed an ablation experiment to analyze and interpret the model.

What I get

This is my first research experience, where I participated in everything from literature review to experimental design to paper writing. During this time, my coding skills improved significantly, with technical aspects including neural network training, model building, data processing, environment setup (such as Docker), and source code reading and modification. A large number of experiments also helped make my code more organized and cleaner, and I learned efficient time management and how to keep experimental records to work efficiently.

At the same time, this experience made me realize that I really enjoy exploring the unknown and providing my own solutions. The thrill of exploring uncharted territories captivated me and marked the beginning of my enjoyment of research.

What next

I’ll do some research about time series data representation, VAE for Financial data generation next semester~ Also I’ll try to improve our pipeline performance by building better Autoencoder architecture.