Paper deep dive
ForecastAgentSearch: Towards a Multi-Expert Agent Search System for Geopolitical Event Forecasting
Miaomiao Cai, He Chang, Yunshan Ma, See-kiong Ng
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 98%
Last extracted: 7/5/2026, 1:35:52 AM
Summary
ForecastAgentSearch is a proposed multi-expert agent search framework designed for complex geopolitical event forecasting. Unlike single-predictor or fixed-ensemble models, it treats specialized expert agents as searchable, rankable, and composable resources. The system follows a three-stage process: (1) task understanding to identify expertise requirements from heterogeneous evidence (historical, news, multimodal), (2) searching and ranking a fine-grained expert agent space based on profiles (region, domain, reliability, cost), and (3) coordinating selected experts to synthesize a final forecast with explanations and uncertainty awareness. The Middle East is identified as a primary testbed for this framework due to its complex regional dynamics.
Entities (5)
Relation Signals (4)
ForecastAgentSearch → addresses → Geopolitical Event Forecasting
confidence 100% · we present ForecastAgentSearch, a preliminary framework that formulates geopolitical event forecasting as a multi-expert agent search problem.
ForecastAgentSearch → utilizes → Expert Agent
confidence 100% · The selected agents provide specialized analyses, which are further coordinated to generate a final forecast
Large Language Models → enables → ForecastAgentSearch
confidence 90% · Recent advances in large language models (LLMs) have created new opportunities for event forecasting.
Expert Agent → specializesin → Middle East
confidence 90% · Middle East event forecasting serves as a meaningful testbed, since it requires reasoning over regional, political, economic, religious, historical, and multimodal factors.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Geopolitical event forecasting is a challenging task, as it requires understanding complex regional contexts, dynamic event signals, and uncertain future outcomes. Recent advances in large language model agents provide new opportunities for building forecasting systems that can reason with diverse sources and expert perspectives. In this paper, we present \textit{ForecastAgentSearch}, a preliminary framework that formulates geopolitical event forecasting as a multi-expert agent search problem. Given a forecasting query, the system first analyzes the task context, then searches and ranks relevant expert agents based on their regional knowledge, domain expertise, reliability, and complementarity. The selected agents provide specialized analyses, which are further coordinated to generate a final forecast with explanations and uncertainty awareness. We discuss the key design challenges of agent profiling, expert retrieval, ranking, and multi-agent coordination, and outline possible evaluation protocols for future development. This work aims to provide an initial step toward searchable and reliable agent-based forecasting systems.
Tags
Links
- Source: https://arxiv.org/abs/2606.31665v1
- Canonical: https://arxiv.org/abs/2606.31665v1
Trouble viewing inline? Open PDF directly →
Full Text
13,764 characters extracted from source content.
Expand or collapse full text
ForecastAgentSearch: Towards a Multi-Expert Agent Search System for Geopolitical Event Forecasting Miaomiao Cai National University of Singapore Singapore, Singapore cmm.hfut@gmail.com He Chang Communication University of China Beijing, China hechangcuc@cuc.edu.cn Yunshan Ma Singapore Management University Singapore, Singapore ysma@smu.edu.sg See-Kiong Ng National University of Singapore Singapore, Singapore seekiong@nus.edu.sg 1 Extended Abstract Geopolitical event forecasting aims to anticipate future political, military, and social developments from historical observations and evolving real-world contexts [4,12,16]. It is important for early warning, risk assessment, policy planning, and strategic decision- making in high-impact scenarios such as international conflicts, diplomatic actions, sanctions, protests, humanitarian crises, and regional instability [5,6,16]. This task is especially challenging in regions such as the Middle East, where future events are often shaped by intertwined factors, including regional power dynamics, political alliances, economic pressure, religious and cultural ten- sions, historical grievances, and multimodal media narratives [8,9]. Therefore, effective forecasting requires not only temporal reason- ing over historical events, but also the ability to identify which evidence and expertise are useful for a specific forecasting task. Recent advances in large language models (LLMs) have cre- ated new opportunities for event forecasting. LLM-based meth- ods can process textual contexts, retrieve relevant historical ev- idence, and generate predictions through prompting, in-context learning, chain-of-thought reasoning, or retrieval-augmented gen- eration [3,7,10,11]. However, most existing approaches still rely on either a single general-purpose predictor or a fixed expert en- semble. A single predictor may follow a dominant reasoning path and overlook alternative geopolitical perspectives, while a fixed expert ensemble may introduce redundant, irrelevant, or costly expert opinions [2,8,15]. Such designs are insufficient for com- plex geopolitical forecasting, where different queries may require different combinations of regional, political, economic, religious, historical, multimodal, and risk-oriented expertise. Recent forecasting studies suggest two important observations. First, heterogeneous evidence sources may play different functional roles in prediction. For example, multimodal evidence can either highlight salient historical events or provide complementary con- text beyond textual descriptions [9]. Second, specialized expert models can provide complementary predictive knowledge, but their usefulness is often query-dependent [2,8]. These observations indicate that reliable geopolitical forecasting requires more than Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from permissions@acm.org. AgentSearch @ SIGIR ’26, Melbourne | Naarm, Australia © 2026 Copyright held by the owner/author(s). Publication rights licensed to ACM. ACM ISBN 978-x-x-x-x/Y/M https://doi.org/10.1145/n.n stronger predictive models or larger context windows. It also re- quires a principled mechanism for deciding which experts should be consulted, how they should be ranked, and how their outputs should be coordinated. To this end, we introduce ForecastAgentSearch, a system-level formulation that treats complex geopolitical event forecasting as a multi-expert agent search problem. Rather than directly mapping historical contexts to predictions, ForecastAgentSearch introduces an intermediate search and coordination layer over specialized expert agents. In this formulation, expert agents are treated as searchable, rankable, and composable resources, which naturally connects geopolitical forecasting with core problems in information retrieval and agent search [1, 13, 14]. As illustrated in Figure 1, ForecastAgentSearch consists of three main stages. The first stage takes a forecasting task and heteroge- neous evidence as input, including the forecasting query, historical events, news context, and multimodal evidence. These sources pro- vide complementary signals for prediction: historical events capture temporal actor interactions, news reports provide contextual narra- tives, and multimodal evidence may reveal additional regional or situational information. Instead of treating all evidence uniformly, ForecastAgentSearch performs task understanding to infer the ex- pertise requirements behind the current query, such as regional locality, political relations, economic pressure, religious or cultural factors, security risks, and historical context. The second stage searches over a fine-grained expert agent space. Each expert agent is associated with a lightweight profile that de- scribes its region or actor specialization, domain expertise, sup- ported evidence sources, reliability estimate, inference cost, and known limitations. For example, a forecasting query may require experts on Israeli political dynamics, Saudi cultural factors, U.S. policy, oil markets, religious tensions, multimodal conflict evidence, or regional security risks. These profiles allow expert agents to be indexed and retrieved according to both structured metadata and semantic descriptions. Given a forecasting query, ForecastAgentSearch retrieves and ranks candidate agents according to several task-aware criteria, in- cluding relevance to the query, regional or actor locality, historical reliability, inference cost, and complementarity with other selected experts. Rather than consulting all available agents, the system selects a compact set of top-ranked experts that can provide com- plementary perspectives for the current task. This design reflects the nature of geopolitical forecasting: the usefulness of an expert is highly dependent on the specific event, region, actors, and evidence sources involved. arXiv:2606.31665v1 [cs.MA] 30 Jun 2026 AgentSearch @ SIGIR ’26, July 24, 2026, Melbourne | Naarm, AustraliaMiaomiao Cai et al. Figure 1: Overview of ForecastAgentSearch. Given a geopolitical forecasting query and heterogeneous evidence, the system first identifies task-specific expertise requirements, then searches over a fine-grained expert agent space, retrieves and ranks the most suitable agents, and coordinates their outputs to generate the final forecast with explanations and uncertainty estimates. The final stage coordinates the selected experts to produce the forecast. Each expert can contribute a prediction, an intermediate analysis, supporting evidence, a confidence estimate, or possible risk factors from its own perspective. A coordination module then synthesizes these outputs into the final forecast, together with ex- planations and uncertainty signals. When experts disagree, the coordinator can compare their evidence, reliability, and domain coverage, rather than simply averaging their predictions. This pro- cess resembles a structured think-tank workflow, where multiple specialists contribute different views and a coordinator synthesizes them into a coherent judgment. The main contribution of this extended abstract is threefold. First, we formulate complex geopolitical event forecasting as an expert- agent search problem, shifting the focus from prediction alone to task-aware expert selection and coordination. Second, we outline ForecastAgentSearch, a system framework that retrieves, ranks, and coordinates specialized agents with different forms of expertise, including regional, political, economic, religious, historical, multi- modal, and risk-oriented knowledge. Third, we discuss Middle East event forecasting as a representative testbed for this formulation, due to its complex regional interactions, heterogeneous evidence sources, and diverse analytical perspectives. Although this extended abstract focuses on the problem formu- lation and system design, ForecastAgentSearch naturally suggests several evaluation directions. Forecasting quality can be measured by accuracy, Brier score, log score, calibration error, temporal gen- eralization, and uncertainty quality. The quality of agent search can be evaluated by determining whether retrieved experts match the required regions, domains, actors, and evidence sources. Future ab- lations can compare task-aware agent search with single-predictor forecasting, fixed expert ensembles, all-expert consultation, ran- dom expert selection, and variants without reliability-, cost-, or complementarity-aware ranking. 2 Conclusion This extended abstract introduces ForecastAgentSearch, which formulates complex geopolitical event forecasting as a multi-expert agent search problem. Instead of relying on a single predictor or a fixed expert ensemble, ForecastAgentSearch treats expert agents as searchable and composable resources, and aims to retrieve, rank, and coordinate suitable experts according to task-specific require- ments. Motivated by recent findings that both evidence utility and expert usefulness are task-dependent, ForecastAgentSearch highlights the need for structured expert profiling, task-aware retrieval, reliability- and cost-aware ranking, complementary team formation, and in- terpretable aggregation. Middle East event forecasting serves as a meaningful testbed, since it requires reasoning over regional, political, economic, religious, historical, and multimodal factors. References [1] Norbert Braunschweiler, Rama Doddipatla, and Tudor-Catalin Zorila. 2025. Tool- ReAGt: tool retrieval for LLM-based complex task solution via retrieval aug- mented generation. In Proceedings of the 3rd Workshop on Towards Knowledgeable Foundation Models (KnowFM). 75–83. [2] Weilin Cai, Juyong Jiang, Fan Wang, Jing Tang, Sunghun Kim, and Jiayi Huang. 2024. A survey on mixture of experts. Authorea Preprints (2024). ForecastAgentSearch: Towards a Multi-Expert Agent Search System for Geopolitical Event Forecasting AgentSearch @ SIGIR ’26, July 24, 2026, Melbourne | Naarm, Australia [3]He Chang, Chenchen Ye, Zhulin Tao, Jie Wu, Zhengmao Yang, Yunshan Ma, Xianglin Huang, and Tat-Seng Chua. 2024. A comprehensive evaluation of large language models on temporal event forecasting. arXiv preprint arXiv:2407.11638 (2024). [4]Songgaojun Deng, Maarten de Rijke, and Yue Ning. 2024. Advances in human event modeling: From graph neural networks to language models. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 6459–6469. [5] Danny Halawi, Fred Zhang, Chen Yueh-Han, and Jacob Steinhardt. 2024. Ap- proaching human-level forecasting with language models. Advances in Neural Information Processing Systems 37 (2024), 50426–50468. [6]Woojeong Jin, Rahul Khanna, Suji Kim, Dong-Ho Lee, Fred Morstatter, Aram Galstyan, and Xiang Ren. 2021. Forecastqa: A question answering challenge for event forecasting with temporal text data. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 4636– 4650. [7] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al.2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in neural information processing systems 33 (2020), 9459–9474. [8] Haoxuan Li, He Chang, Yunshan Ma, Yi Bin, Yang Yang, See-Kiong Ng, and Tat-Seng Chua. 2026. ThinkTank-ME: A Multi-Expert Framework for Middle East Event Forecasting. In W. [9]Haoxuan Li, Zhengmao Yang, Yunshan Ma, Yi Bin, Yang Yang, and Tat-Seng Chua. 2024. M-Forecast: A Multimodal Approach to Temporal Event Forecasting with Large Language Models. In M. [10]Ruotong Liao, Xu Jia, Yangzhe Li, Yunpu Ma, and Volker Tresp. 2024. Gentkg: Generative forecasting on temporal knowledge graph with large language models. In Findings of the association for computational linguistics: NAACL 2024. 4303– 4317. [11]Ruilin Luo, Tianle Gu, Haoling Li, Junzhe Li, Zicheng Lin, Jiayi Li, and Yujiu Yang. 2024. Chain of history: Learning and forecasting with llms for temporal knowledge graph completion. arXiv preprint arXiv:2401.06072 (2024). [12] Yunshan Ma, Chenchen Ye, Zijian Wu, Xiang Wang, Yixin Cao, and Tat-Seng Chua. 2023. Context-aware event forecasting via graph disentanglement. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 1643–1652. [13] Zhengliang Shi, Yuhan Wang, Lingyong Yan, Pengjie Ren, Shuaiqiang Wang, Dawei Yin, and Zhaochun Ren. 2025. Retrieval models aren’t tool-savvy: Bench- marking tool retrieval for large language models. In Findings of the Association for Computational Linguistics: ACL 2025. 24497–24524. [14]Bin Wu, Arastun Mammadli, Xiaoyu Zhang, and Emine Yilmaz. 2026. AgentSearchBench: A Benchmark for AI Agent Search in the Wild. arXiv preprint arXiv:2604.22436 (2026). [15]Chenchen Ye, Ziniu Hu, Yihe Deng, Zijie Huang, Mingyu Derek Ma, Yanqiao Zhu, and Wei Wang. 2024. Mirai: Evaluating llm agents for event forecasting. arXiv preprint arXiv:2407.01231 (2024). [16]Liang Zhao. 2021. Event Prediction in the Big Data Era: A Systematic Survey. Comput. Surveys 54, 5 (2021), 1–37.