LOBBen-TM: A Benchmark Study of Limit Order Book Prediction with Temporal Modeling

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

Keywords: limited order book, time series prediction, mid-price trend prediction, mid-price return forecasting, benchmark, open source datasets

Abstract: We introduce LOBBen-TM, a limit order book (LOB) benchmark with temporal modeling for deep learning on open-sourced LOB data that unifies evaluations across tasks, features, and assets. Our work makes four major contributions: (i) On the Mid-Price Trend Prediction (MPTP) task, we assess state-of-the-art LOB models with a standardized full LOB feature set with time-sensitive features on two assets, equities (FI-2010) and cryptocurrency (Bitcoin), to probe cross-asset generalization under a common protocol. We further benchmark a common LOB feature taxonomy (basic, time-insensitive, time-sensitive) and conduct an ablation on FI-2010. (ii) We extend the study to Mid-Price Return Forecasting (MPRF), jointly evaluating LOB-specific architectures and top-tier general time-series predictors on FI-2010 with MSE, R2, and Pearson correlation. (iii) To enhance multivariate time series prediction models on LOB returns, we propose a lightweight Cross-Variate Mixing Layer (CVML) that plugs into existing models. Empirically, results show that the standardized full feature set yields robust MPTP performance across FI-2010 and Bitcoin while revealing asset-dependent ranking shifts. Besides, time-sensitive features provide sizable improvements on FI-2010, underscoring the importance of temporal signal modeling. Last but not least, our proposed CVML architecture substantially boosts general time series prediction models on MPRF, narrowing the gap to LOB models and advancing return forecasting on LOB data.

Primary Area: datasets and benchmarks

Submission Number: 649