The paper "Learning to Generate Explainable Stock Predictions using Self-Reflective Large Language Models" addresses the challenge of explaining stock predictions, which is traditionally difficult for non-generative deep learning models. The authors propose the Summarize-Explain-Predict (SEP) framework, which uses a verbal self-reflective agent and Proximal Policy Optimization (PPO) to enable a Large Language Model (LLM) to generate explainable stock predictions in an autonomous manner. The SEP framework consists of three main components: Summarize, Explain, and Predict. The Summarize module converts large volumes of text into point-form summaries, the Explain module teaches the LLM to generate correct stock predictions and explanations through a self-reflective process, and the Predict module fine-tunes the LLM using its self-generated annotated samples. The authors demonstrate that their fine-tuned LLM outperforms both traditional deep-learning and LLM methods in prediction accuracy and Matthews correlation coefficient (MCC) for stock classification. They also validate the framework's generalizability by fine-tuning the LLM to generate quantitative weights for multiple stocks in a portfolio task, showing improved profitability and Sharpe Ratio. The paper discusses the ethical implications of using LLMs for stock prediction and suggests measures to mitigate potential risks.The paper "Learning to Generate Explainable Stock Predictions using Self-Reflective Large Language Models" addresses the challenge of explaining stock predictions, which is traditionally difficult for non-generative deep learning models. The authors propose the Summarize-Explain-Predict (SEP) framework, which uses a verbal self-reflective agent and Proximal Policy Optimization (PPO) to enable a Large Language Model (LLM) to generate explainable stock predictions in an autonomous manner. The SEP framework consists of three main components: Summarize, Explain, and Predict. The Summarize module converts large volumes of text into point-form summaries, the Explain module teaches the LLM to generate correct stock predictions and explanations through a self-reflective process, and the Predict module fine-tunes the LLM using its self-generated annotated samples. The authors demonstrate that their fine-tuned LLM outperforms both traditional deep-learning and LLM methods in prediction accuracy and Matthews correlation coefficient (MCC) for stock classification. They also validate the framework's generalizability by fine-tuning the LLM to generate quantitative weights for multiple stocks in a portfolio task, showing improved profitability and Sharpe Ratio. The paper discusses the ethical implications of using LLMs for stock prediction and suggests measures to mitigate potential risks.