This paper investigates whether increasing model complexity in deep learning architectures genuinely improves short-term price prediction from limit order book (LOB) data, or if gains are primarily driven by data preprocessing and feature engineering. Using 100ms snapshots of BTC/USDT from Bybit, the authors benchmark six models—ranging from logistic regression and XGBoost to deep architectures like DeepLOB and CNN+LSTM—on binary and ternary classification tasks at horizons of 100ms, 500ms, and 1 second. Two denoising techniques, Savitzky–Golay smoothing and Kalman filtering, are applied to raw LOB data to mitigate microstructural noise.

The results show that Savitzky–Golay filtering consistently improves predictive accuracy across all models, while Kalman filtering often underperforms relative to raw data. Surprisingly, simpler models such as logistic regression and XGBoost achieve comparable or even superior F1 scores compared to complex neural networks, especially when combined with proper preprocessing. For example, at the 500ms horizon with 40-level LOB data, logistic regression with Savitzky–Golay filtering reaches 72.84% binary accuracy, outperforming DeepLOB (71.89%). The study also reveals a trade-off between LOB depth and data coverage: deeper levels provide more information but reduce the number of usable snapshots, lowering overall accuracy.

The key contribution is demonstrating that careful data preprocessing and feature engineering can be more impactful than adding hidden layers or parameters. This finding has practical implications for high-frequency trading, where model interpretability, low latency, and robustness are critical. The authors advocate for a focus on input quality over architectural complexity in LOB-based forecasting.