Two Heads are Better than One: Distilling Large Language Model Features Into Small Models with Feature Decomposition and Mixture

Overview

The paper presents Cooperative Market Making (CMM), a framework designed to distill features from large language models into smaller, more efficient models for financial trading. This approach addresses the slow inference speed of large models by using feature decomposition and mixture.

Key Insights

• Insight: CMM decouples LLM features across layer, task, and data dimensions.
  • Evidence: Extensive experiments on four real-world market datasets.
  • Verifiable: Yes
• Insight: CMM outperforms current distillation methods and RL-based strategies.
  • Evidence: Experimental results demonstrate superiority.
  • Verifiable: Yes

BFSI Relevance

• Why Relevant: CMM offers a method to enhance trading strategies by using smaller, efficient models.
• Primary Sector: Financial Services
• Subsectors: Asset Management, Trading
• Actionable Implications:
  • Implement CMM to optimize trading algorithms.
  • Explore feature decomposition for model efficiency.
  • Consider integrating CMM in existing trading platforms.