Evaluating the effectiveness of crowd wisdom and large language models for fantasy cricket team selection
研究了群体智慧和推理型大型语言模型在印度超级联赛梦幻板球队选择中的表现，发现规则引导的语言模型能媲美群体智慧，为混合人机决策提供了可扩展框架。
This study evaluates the effectiveness of the Wisdom of Crowds (WOC) and reasoning large language models (LLMs) in constructing high-performing fantasy cricket teams within the constrained decision space of the Indian Premier League (IPL). Fantasy team selection presents a high-dimensional, combinatorial optimisation problem subject to budget, role, and captaincy constraints, making it a natural testbed for operational research and collective intelligence. Using data from over 500 million Dream11 team entries across 40 IPL contests in 2023 and 2025, we perform prediction and evaluation, assessing whether aggregated crowd selections outperform individual users and expert-curated line-ups. Two WOC aggregation strategies, mode-based and greedy role-wise heuristics are benchmarked against population and expert baselines. To interpret WOC behaviour, we extract structural heuristics from top-performing entries, including role configurations, credit mixes, and context-aware co-selection rules using rule mining techniques. These heuristics are then embedded into structured prompts that guide a reasoning LLM to generate teams under operational constraints for the IPL 2025 final. The rule-based LLM outperforms its baseline counterpart and rivals WOC teams in performance, showcasing a scalable framework for hybrid human-AI decision-making. We release anonymised fantasy team datasets to support future work in crowd analytics and sports modelling for decision support.