TRADES diffusion model outperforms LOB

A new line of market-simulation research is getting attention because it targets one of the hardest problems in quantitative finance: generating realistic limit order book data for testing strategies before they touch live markets. (arxiv.org) Leonardo Berti, Bardh Prenkaj and Paola Velardi described that effort in a paper, “TRADES: Generating Realistic Market Simulations with Diffusion Models,” posted to arXiv on January 31, 2025. (arxiv.org) The paper presents TRADES, short for TRAnsformer-based Denoising Diffusion Probabilistic Engine for LOB Simulations, as a model built to generate realistic and responsive order-book simulations. (github.com) The authors also released DeepMarket, an open-source Python framework, along with a synthetic dataset derived from the model. ### Why are researchers focused on limit order book simulation in the first place? Limit order books sit at the core of electronic markets because they record buy and sell orders, queue dynamics and changes in market depth at high frequency. The TRADES paper says realistic LOB simulation matters for calibrating and testing trading strategies, running market-impact experiments and generating synthetic market data. The authors argue previous approaches fell short on realism, usefulness and responsiveness. A synthetic simulator matters because real high-frequency market data is expensive, sensitive and difficult to use for repeated experiments. (arxiv.org) The DeepMarket repository says the framework is designed for LOB simulation with deep learning and includes preprocessing for high-frequency data, model training tools and evaluation components. ### What is TRADES actually doing differently? TRADES uses a transformer-based denoising diffusion approach rather than relying only on more conventional autoregressive generation. In the paper, the authors say the model is designed to learn the conditional distribution of market data and generate simulations that remain responsive to an experimental agent. (arxiv.org) That matters for researchers who want a simulator that does more than replay statistical averages. The paper’s framing is that market simulation has to capture both realism and interaction. The authors write that TRADES can react to an experimental agent and can support calibration and evaluation of trading strategies and market-impact experiments. (github.com) That is a more demanding test than simply producing plausible-looking price paths. ### How strong was the reported performance? The headline result in the paper is a 3.27 to 3.48 times improvement in predictive-score results over prior baselines, according to summaries of the arXiv manuscript and related paper pages. (arxiv.org) Those predictive scores are used as one way to judge how useful the synthetic data is relative to earlier methods. ArXiv’s abstract and mirrored summaries also describe the model as state of the art for generating realistic and responsive market simulations. The paper combines those benchmark-style results with tests of realism and responsiveness rather than reporting only one metric. (arxiv.org) ### What did the authors release alongside the paper? The GitHub repository says DeepMarket is the official codebase for the paper and describes it as the first open-source Python framework for deep-learning-based market simulation. The paper’s repository materials say it includes preprocessing pipelines, a PyTorch Lightning training environment and tools for hyperparameter search. (arxiv.org) The authors also included a synthetic dataset called TRADES-LOB, described in the paper as a dataset composed of simulations generated by TRADES. The paper says that dataset is part of the repository and is intended to support downstream work in market simulation research. (arxiv.org) ### Why does the open-source release matter for quant researchers? DeepMarket gives researchers a common framework for building and comparing simulators instead of stitching together private pipelines. The repository says the package is built for LOB simulation with deep learning, while the paper says the synthetic dataset can be used for calibration, evaluation and strategy-related experiments. (github.com) For researchers working on backtesting or reinforcement-learning agents, the practical value is that the code and synthetic data are already public in the DeepMarket repository maintained under LeonardoBerti00 on GitHub. (arxiv.org) The arXiv paper remains available as versioned research documentation, and the repository is the main place to inspect the framework and dataset files. (arxiv.org) (github.com)