Hot events refer to events that attract intense public attention, trigger extensive discussions, and exert significant influence on society, the economy, or organizations. Early warning of such events can effectively reduce risk losses. Hot event early warning is an issue of complex scenario analysis involving multiple parties, factors, and relationships. Traditional single “indicator-symptom” correlation evaluation systems have problems such as low richness of indicator systems and rigid reasoning processes, thereby easily causing the omission of important symptom clues in complex scenarios and untimely early warnings of hot events. To this end, this study proposes a hot event early warning algorithm based on the dynamic indicator network and large language model. First, a hierarchical and classified indicator network is constructed from a multi-group perspective, with the coupling relationships between indicator systems of different groups established. By conducting indicator enhancement and intensity segmentation, indicators across and within different groups are graded and augmented to improve the efficiency of symptom discovery and event early warning. Second, a hot event early warning method with collaboration between large and small models is designed. LLM is adopted to intelligently schedule two types of models, including classic algorithms and business data models, and the indicator network is combined to complete hot event early warning. The early warning results are adaptively fed back through prompt engineering to generate new indicators and optimize indicator weights. Comparative experimental results show that this method can complete hot event early warning in complex scenarios, and the comprehensiveness of symptom coverage and the timeliness of event early warning are improved, thus verifying the effectiveness of the method.