A Benchmark Study For Limit Order Book (LOB) Models and Time Series Forecasting Models on LOB Data

Weijian Li, Stephen S Cheng, Lining Mao, Jigyasa Kumari, Alex Pyo, Mehak Kawatra, Jialong Li, Jiayi Wang, Ammar Gilani, Jingya Xun, Jerry Yao-Chieh Hu, Han Liu

Keywords: benchmark, time series forecasting, convolution, deep learning, limit order book, mid-price trend prediction, mid-price return forecasting

Abstract: We present a comprehensive benchmark to evaluate the performance of deep learning models on limit order book (LOB) data. Our work makes four significant contributions: (i) We evaluate existing LOB models on a proprietary futures LOB dataset to examine the transferability of LOB model performance between various assets; (ii) We are the first to benchmark existing LOB models on the mid-price return forecasting (MPRF) task. (iii) We present the first benchmark study to evaluate SOTA time series forecasting models on the MPRF task to bridge the two fields of general-purpose time series forecasting and LOB time series forecasting; and (iv) we propose an architecture of convolutional cross-variate mixing layers (CVML) as an add-on to any deep learning multivariate time series model to significantly enhance MPRF performance on LOB data. Our empirical results highlight the value of our benchmark results on our proprietary futures LOB dataset, demonstrating a performance gap between the commonly used open-source stock LOB dataset and our futures dataset. Furthermore, the results demonstrate that LOB-aware model design is essential for achieving optimal prediction performance on LOB datasets. Most importantly, our results show that our proposed CVML architecture brings about an average improvement of 244.9% to various time series models’ mid-price return forecasting performance.

Primary Area: learning on time series and dynamical systems

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 2746