@@html:<div class=”row”><h2 class=”col-md-12 text-center”><strong><font size=”+4r”>@@ LOB-Bench: Benchmarking Generative AI for Finance – an Application to Limit Order Book Data @@html:</font></strong></h2></div>@@
- Peer Nagy*
- Sascha Frey*
- Kang Li
- Bidipta Sarkar
- Svitlana Vyetrenko
- Stefan Zohren
- Anisoara Calinescu
- Jakob Foerster
- ICML 2025
- @@html:<a href=”https://arxiv.org/abs/2502.09172”><image src=”imgs/paper_front.jpg” height=”60px”><h4><strong>Paper</strong></h4></a>@@
- @@html:<a href=”https://github.com/peernagy/lob_bench”><image src=”imgs/GitHub-Mark.png” height=”60px”><h4><strong>Code</strong></h4></a>@@
- @@html:<a href=”https://huggingface.co/datasets/peernagy/lob_bench”><image src=”imgs/huggingface.svg” height=”60px”><h4><strong>Dataset</strong></h4></a>@@
While financial data presents one of the most challenging and interesting sequence modelling tasks due to high noise, heavy tails, and strategic interactions, progress in this area has been hindered by the lack of consensus on quantitative evaluation paradigms. To address this, we present LOB-Bench, a benchmark, implemented in python, designed to evaluate the quality and realism of generative message-by-order data for limit order books (LOB) in the LOBSTER format. Our framework measures distributional differences in conditional and unconditional statistics between generated and real LOB data, supporting flexible multivariate statistical evaluation. The benchmark also includes features commonly used LOB statistics such as spread, order book volumes, order imbalance, and message inter-arrival times, along with scores from a trained discriminator network. Lastly, LOB-Bench contains “market impact metrics”, i.e. the cross-correlations and price response functions for specific events in the data. We benchmark generative autoregressive state-space models, a (C)GAN, as well as a parametric LOB model and find that the autoregressive GenAI approach beats traditional model classes.
Please refer to the benchmark readme for information on how to use along with the tutorial notebook.
We evaluate models on a subsample of data from January 2023 (except for Coletta which was trained and tested on January 2019). If you would like to add your model to this leaderboard, please contact the authors via email, and provide the directory of CSVs with data_real, data_gen, and data_cond as shown in the tutorial notebook so we can reproduce your results. We highly encourage open-source development of models, and will provide links to your code or model if provided.
| Model | Wasserstein (↓) | L1 (↓) | Link |
|---|---|---|---|
| LobS5 [cite:@nagy2025lobbench;@nagy2023generative] | 0.16 | 0.14 | |
| Baseline [cite:@nagy2025lobbench;@cont2010stochastic] | 0.29 | 0.36 | |
| RWKV4 [cite:@nagy2025lobbench;@peng2023rwkv] | 0.36 | 0.29 | |
| RWKV6 [cite:@nagy2025lobbench;@peng2024eagle] | 0.45 | 0.31 | |
| Coletta [cite:@nagy2025lobbench;@coletta2023conditional] | 0.54 | 0.48 |
| Model | Wasserstein (↓) | L1 (↓) | Link |
|---|---|---|---|
| LobS5 [cite:@nagy2025lobbench;@nagy2023generative] | 0.19 | 0.13 | |
| RWKV6 [cite:@nagy2025lobbench;@peng2024eagle] | 0.26 | 0.32 | |
| Baseline [cite:@nagy2025lobbench;@cont2010stochastic] | 0.69 | 0.5 | |
| RWKV4 [cite:@nagy2025lobbench;@peng2023rwkv] | 0.81 | 0.61 |