Aligning Language Models with Investor and Market Behavior for Financial Recommendations

Battle Count: 136

The paper is moderately relevant to quantitative trading. It focuses on asset recommendation (which assets to buy/sell) rather than direct trading strategy optimization or execution. The behavioral alignment component is more relevant to retail wealth management than institutional quant trading. However, the profitability metric (Prof@3) and the federated learning approach for cross-institutional collaboration could be adapted for trading signal generation. The KG-based market trend encoding and 180-day horizon are more aligned with medium-term portfolio management than high-frequency or systematic trading.

Moderate-to-high complexity. Requires: (1) KG construction from transaction and price data with JSON-LD serialization, (2) LLM fine-tuning with KTO (binary labels), (3) LoRA + 4-bit quantization setup, (4) YaRN context extension to 131K tokens, (5) Federated learning infrastructure with client simulation, (6) Multiple model size evaluations. The centralized version is more accessible, while the federated variant requires distributed training orchestration. The KG construction pipeline and prompt engineering add domain-specific complexity.

GitHub repository with source code, dataset construction scripts, and result generation pipelines is available. FAR-Trans dataset is referenced. LoRA configuration (rank 16, alpha 64), 4-bit quantization, and training protocol (3 epochs centralized, 200 rounds federated with 3 clients per round) are specified. Qwen3 models are publicly available. However, exact random seeds and full hyperparameter sweep details are not fully enumerated.