Learning to Generate Explainable Stock Predictions using Self-Reflective Large Language Models