The paper introduces FinLlama, a novel financial sentiment analysis framework designed specifically for algorithmic trading applications. The approach addresses limitations of existing methods by developing a finance-specific Large Language Model (LLM) based on the Llama 2 7B foundational model. Unlike standard lexicon-based approaches that suffer from context sensitivity issues, or general-purpose LLMs that lack financial specificity and require substantial computational resources, FinLlama combines the generative capabilities of Llama 2 with a neural network-based decision mechanism to create a sophisticated generator-classifier scheme.

The methodology involves fine-tuning the Llama 2 7B model on supervised financial sentiment data while implementing parameter-efficient fine-tuning through LoRA (Low-Rank Adaptation) to optimize trainable parameters and minimize computational requirements. This approach enables FinLlama to not only classify sentiment valence but also quantify its strength, providing traders with nuanced insights into financial news articles. Simulation results demonstrate that FinLlama significantly enhances portfolio management decisions, generating higher returns while maintaining resilience during volatile market periods and unpredictable events, thereby establishing its practical value for algorithmic trading applications.