News-Aware Direct Reinforcement Trading for Financial Markets

Battle Count: 88

Highly relevant to quantitative trading as it demonstrates a practical end-to-end RL framework that eliminates handcrafted features, directly incorporating LLM-derived news sentiment with raw market data. The results show consistent outperformance over BTC market benchmarks (124.5% vs 56% baseline). The ablation studies clearly demonstrate the value of both news signals and sequence-based architectures. However, the simplified action space and single-asset focus limit immediate practical applicability.

Moderate complexity. The pipeline involves: (1) LLM API calls for sentiment extraction with batch processing, (2) time-series data preprocessing and merging, (3) RL agent training with DDQN/GRPO, (4) sequence model architectures (LSTM/Transformer), and (5) extensive hyperparameter tuning via Optuna with TPE. The use of standard RL algorithms and well-known architectures reduces some complexity, but the end-to-end integration and hyperparameter sensitivity of RL add significant engineering challenges.

The paper provides detailed hyperparameter ranges (Table 3), dataset sources (Binance for OHLCV, HuggingFace for Bitcoin news), and algorithm descriptions. However, no code repository is mentioned, and the LLM prompt template is shown only as a figure. The use of Gemini-2.5-flash API introduces non-determinism. The 70/15/15 chronological split and evaluation protocol are clearly described.