When Less Is More: Domain-Aware Dual-Branch Recurrent Networks for Limit Order Book Mid-Price Prediction

Abstract

Dual-branch BiGRU-CNN for high-frequency LOB mid-price prediction. Weighted Pearson rho_w = 0.266, outperforming LightGBM on 58.3% of sequences. Documents feature sufficiency and the negative ensemble effect on HF data (Wunder Fund RNN challenge).

Key questions

Q. Can deep learning predict limit-order-book mid-price movements?
A domain-aware dual-branch BiGRU-CNN reached a weighted Pearson ρ_w = 0.266 and outperformed a LightGBM baseline on 58.3% of sequences (Wunder Fund RNN challenge).

Q. Do larger or ensembled models always help on high-frequency financial data?
No. The study documents a negative ensemble effect on high-frequency LOB data — combining models hurt performance — and emphasizes feature sufficiency over model size (“when less is more”).