PandaAI: A Practical Agent CQ2 for Neuro-symbolic Data Analysis And Integrated Decision-Making in Quantitative Finance

Battle Count: 95

Directly addresses core quantitative trading challenges: alpha factor discovery, portfolio optimization, risk management, and execution. The paper explicitly models the sequential decision-making process for portfolio weights, incorporates realistic transaction costs, market frictions, and turnover constraints. The neuro-symbolic approach bridges LLM reasoning with financial rigor, and the closed-loop system mirrors the actual quant research lifecycle. The evaluation uses industry-standard metrics (IC, Rank IC, ICIR, AR, MDD) on a major equity index.

Extremely complex multi-module system requiring: (1) Fine-tuning a 33B parameter LLM with SFT + RLHF pipeline, (2) Training an autoencoder for latent regime states, (3) Implementing dual-channel adapters (MLP for soft tokens + feature extraction network for scalars), (4) Building a full MCTS framework with constraint checking, (5) Implementing convex portfolio optimization with regime-conditioned parameters, (6) Designing execution control logic, (7) Building the closed-loop update system with fast symbolic induction and slow LoRA adaptation, (8) Training a reward model for RLHF. Requires significant GPU resources, financial domain expertise, and sophisticated software engineering.

Fine-tuning dataset is explicitly stated as not publicly available due to privacy obligations and NDAs. The system architecture is described in detail with pseudocode (Algorithm 1), but the proprietary nature of the training data, the specific Barra factor implementations, and the closed-loop update mechanisms make full reproduction difficult. No code or model weights are released. The base model (DeepSeek-Coder-33B) is open-source, but the fine-tuning pipeline and constraint sets are not provided.