One More Question is Enough, Expert Question Decomposition (EQD) Model for Domain Quantitative Reasoning

Battle Count: 80

The paper addresses financial quantitative reasoning and QA, which is tangentially related to quantitative trading. It improves LLM performance on financial document analysis and numerical reasoning tasks relevant to fundamental analysis. However, it does not directly address trading strategies, market prediction, portfolio optimization, or risk management. The financial QA capability could support research and analysis workflows in quantitative finance but is not a trading system itself.

Requires two-step fine-tuning (instruction tuning + PPO reinforcement learning), LoRA adapters, and a reward function implementation. However, resource requirements are modest: single A100 GPU, ~3000 training examples, 22M adapter parameters (0.27% of base model). Training time: ~2.5 hours (step 1) + ~4 hours (step 2). The PPO implementation with multiple model roles (QA, QD, reference) managed via adapter activation adds moderate complexity. Inference is simple: generate 1-2 sub-questions and append to QA prompt.

Code is publicly available on GitHub. Training datasets (ConvFinQA, FinQA) are publicly available. Detailed implementation parameters provided in appendices (LoRA rank=8, alpha=16, batch sizes, learning rates, iterations). Evaluation code released. However, some closed-source API costs and specific checkpoint selection criteria add minor complexity.