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

Abstract

Predicting short-term mid-price movements from limit order book (LOB) data is a fundamental problem in quantitative finance and market microstructure research, with direct applicability to both traditional exchanges and cryptocurrency markets—including centralized exchanges (CEXs) and emerging on-chain LOB protocols in decentralized finance (DeFi). We present three contributions to this domain. First, we propose DA-BiGRU-CNN, a domain-aware dual-branch architecture that decomposes LOB features into price and volume information channels, processes them through dedicated bidirectional GRU encoders with shared microstructure features, and fuses temporal representations via a multi-scale convolutional bottleneck (Conv1d with kernels k = 3,5,7). Second, we provide empirical evidence for a "feature sufficiency" hypothesis: a unidirectional GRU trained on 53 basic features achieves performance statistically equivalent to one trained on 219 extensively engineered features—including rolling statistics, exponential moving averages, and lag/difference features—suggesting that recurrent hidden states implicitly learn these temporal patterns. Third, we document a "negative ensemble effect" where combining sequential (GRU) and tabular (gradient boosting) models consistently degrades prediction quality, contradicting the widely-held assumption that model diversity improves ensemble performance. On a large-scale dataset of 12,165 LOB sequences (12.1M timesteps), our GRU baseline achieves a weighted Pearson correlation of 0.266, outperforming LightGBM by 58%, while our domain-aware architecture offers an architecturally principled alternative that naturally separates price dynamics from liquidity dynamics.