FinSMART: Financial Sentiment Analysis for Algorithmic Trading through Market-Aligned Reinforcement Learning

Battle Count: 117

Highly relevant to quantitative trading. The paper directly addresses the generation of trading signals from financial text, evaluates performance through realistic portfolio construction (35% long-short, next-day open-to-open returns), and demonstrates substantial improvements in profitability (264.9% cumulative return vs 109.8% for FinDPO) and risk-adjusted metrics. The continual retraining capability is particularly relevant for live trading systems. The framework eliminates dependence on static labeled datasets, enabling adaptive signal generation in evolving markets.

Moderate complexity. Requires understanding of GRPO/RL post-training, LLM fine-tuning with LoRA, NER pipelines, and financial data processing. However, the authors emphasize computational efficiency (single A6000 GPU, 8-hour training, 13.6M trainable parameters). The three-stage pipeline (data alignment, reward computation, GRPO optimization) is well-structured. Key challenges include proper NER filtering, reward function calibration, and ensuring no look-ahead bias in evaluation.

The paper provides detailed descriptions of the GRPO training configuration (G=8, beta=0.1, LoRA rank=16, alpha=32, dropout=0.05), data sources (The Motley Fool, MarketWatch, Yahoo Finance), and evaluation framework. However, no code repository is mentioned, and the specific hyperparameters for the reward threshold (tau=0.5%) and training schedule are provided. The base model (Llama-3-8B-Instruct) is publicly available. Training completed in 8 hours on a single NVIDIA A6000 GPU.