RESEARCH ARTICLE

Playing the Odds: Agentic LLMs for Real-Time NBA Forecasting and Market Betting

Ethan Jeon — Yongsan International School of Seoul, 285 Itaewon-ro, Yongsan-gu, Seoul 04347, Republic of Korea

Minseong (Leo) Sim — Chadwick International School, 45 Art Center-daero 97beon-gil, Yeonsu-gu, Incheon 21985, Republic of Korea

Woohyun Kim — Seoul Science High School, 63 Hyehwa-ro, Jongno-gu, Seoul 03066, Republic of Korea

10.ethan.jeon@gmail.com

Abstract

Predicting the outcomes of professional basketball games is a challenging problem due to the intrinsic stochastic uncertainty of sports competitions and the heterogeneity of relevant information sources. While existing approaches in sports analytics primar ily rely on structured historical statistics, such methods often struggle to incorporate timely and unstructured information. In this paper, we propose a unified framework that leverages large language models for probabilistic forecasting and decision making for NBA games and their prediction markets. Our approach combines specialized information retrieval agents with multiple role-based LLM predictors, whose forecasts are aggregated into the final forecasting probabilities. These probabilities are then operationalized through a fractional Kelly betting strategy in binary prediction markets. We evaluate the proposed system using both Brier Scores and simulated market returns, demonstrating that LLM -based forecasting can effectively complement traditional models and translate predictive improvements into economic values. Overall, our results demonstrate the potential of LLMs and in-context learning as flexible tools for decision-oriented sports analytics.

Keywords

Keywords: Sports analytics, large language models, Game prediction

Introduction

Advances in foundation models enable new forecasting pipelines, but sports domains demand careful calibration and robustness checks. Agentic designs can improve modularity by separating data retrieval from reasoning and verification.

Conclusion

Agentic LLM approaches may improve structured forecasting workflows, but responsible deployment requires transparency, uncertainty reporting, and adherence to applicable regulations. Further empirical validation on historical seasons is needed to quantify edge and stability.

Download PDF

How to Cite

Jeon, Ethan; Sim, Minseong; Kim, Woohyun. Playing the Odds: Agentic LLMs for Real-Time NBA Forecasting and Market Betting. Journal of Youth Impact. April 2026; 1(Issue 2). DOI: https://doi.org/10.66245/jyi.v1.i2.007