Improving Zero-shot LLM Re-Ranker with Risk Minimization