LLM-SoccerArena: Benchmarking LLMs on Real-World Predictions in Sports
The paper tests whether LLM forecasts improve when they are made live, before match outcomes are known.
LLM-SoccerArena records timestamped predictions for unresolved sports events, along with prompts, model versions, tool traces, and costs. The authors demonstrate it on the 2026 FIFA World Cup, with seven LLMs forecasting all 104 matches and 15 tournament questions. The benchmark varies model version, information access, prompting strategy, and forecast horizon. Web access helped, but only slightly: a 0.023 improvement in Brier score. ArXiv · AI/CL/LG's note
LLM-SoccerArena records timestamped predictions for unresolved sports events, along with prompts, model versions, tool traces, and costs. The authors demonstrate it on the 2026 FIFA World Cup, with seven LLMs forecasting all 104 matches and 15 tournament questions. The benchmark varies model version, information access, prompting strategy, and forecast horizon. Web access helped, but only slightly: a 0.023 improvement in Brier score. ArXiv · AI/CL/LG's note
score 4