Event History Over Scale: Compact Transformers for Low-Latency Limit Order Book Forecasting

Battle Count: 50

Highly relevant for high-frequency trading (HFT) and market making strategies where low latency and small model footprint are critical constraints. Demonstrates that fine-grained event data can outperform aggregated data with significantly smaller models.

The architecture is relatively simple (small transformer with few parameters). The main complexity lies in the data preprocessing pipeline to convert raw L3 message streams into the specific event windows and features required by the model.

Code is available on GitHub. The methodology is clearly described, including hyperparameters and data splits. However, the EPEX electricity data is noted as not publicly available, which limits full reproducibility for that specific dataset.