NTU team’s AI tool among 14 projects worldwide funded by OpenAI

Hyeokkoo Eric Kwon (second from left) with his teammates (from left) Shin Kayoung, Mo Jiayun and Park Jaecheol at Nanyang Business School on Sept 1. The team developed a tool that forecasts economic outcomes, such as whether people will save or spend their money, with far greater accuracy than traditional simulation models.

Published Sep 13, 2026, 05:00 AM

Updated Sep 13, 2026, 05:00 AM

SINGAPORE – Researchers in Singapore are building an artificial intelligence tool that can predict how economic policies such as cash handouts may boost the economy.

The tool forecasts economic outcomes, such as whether people will save or spend their money, with far greater accuracy than traditional simulation models. This allows policymakers to test more alternative policies to achieve the desired economic outcome.

Developed by four researchers at Nanyang Technological University, it is among 14 projects worldwide to receive funding from OpenAI as part of the company’s commitment to promoting economic opportunity and societal resilience through AI.

The 14 projects were selected from over 400 submissions to OpenAI’s call for proposals in April.

The NTU team, led by Nanyang Business School Provost’s Chair Professor in Information Systems Hyeokkoo Eric Kwon , will receive US$100,000 (S$126,800).

He said the money will go towards developing and testing the tool, which has not been named, over the next six months.

The NTU team’s AI prediction system is trained on real transaction data from more than one million users of a South Korean mobile budgeting app, which Kwon declined to name.

The data contains information on how much users spend, which retailers they spend at , and how much monthly income they receive d between 2023 and 2025.

The app obtains the data by tracking notifications, messages and e-mails that are sent to users by banks when transactions are made. Names and bank account numbers are removed. Spending amounts, income and purchase times are also grouped into broader ranges.

The NTU team’s AI tool ingests only the anonymised data to prevent the identification of users .

“There should never be a situation where the data can only be matched to one profile. Depending on the level of sensitivity of the data, a profile is sufficiently anonymised if a minimum of three, five or 10 users share those exact same characteristics,” said Kwon.

AI agents are then used to simulate how households may respond to a proposed government policy.

For example, if the government is planning to give $100 in cash vouchers to every citizen, the system aims to estimate how much additional spending it could generate, what it would be spent on, and how the outcome would change with a different policy, said Kwon.

As AI understands complex patterns and has a vast knowledge of human behaviour, it can predict human behaviour better than traditional simulations based on deterministic rules, he said.

“Qualitative interviews and surveys are conducted to set these rules, but there is often a gap between how people actually behave and how they respond in surveys,” said Kwon, adding that the complex relationship between factors such as income and age cannot be represented in traditional simulation models.

To ensure that the system does not hallucinate and can accurately predict real-world outcomes, it will be validated against real spending data in 2025.

This data was recorded before and after the South Korean government distributed two tranches of cash handouts in July and September 2025. These handouts had to be spent at eligible retailers by November 2025.

Half of the NTU team’s US$100,000 funding will be in cash, while the other half will be in credits for OpenAI’s AI model subscription plans.

By February 2027, the NTU researchers aim to have the tool work for any population.

The tool, which will be made open-source, will also be able to analyse sentiments such as interests and political views to improve prediction accuracy and explain how it arrives at its conclusion.

AI/artificial intelligence

Technology and research

9月1日，Hyeokkoo Eric Kwon（左二）与他的队友（从左至右）Shin Kayoung、Mo Jiayun和Park Jaecheol在南洋商学院。该团队开发了一种工具，可以比传统的模拟模型更准确地预测经济结果，例如人们是会储蓄还是会消费。

发布于 2026 年 9 月 13 日上午 5:00

更新于2026年9月13日凌晨5:00

新加坡——新加坡的研究人员正在开发一种人工智能工具，该工具可以预测现金补贴等经济政策将如何促进经济增长。

该工具能够比传统模拟模型更准确地预测经济结果，例如人们是会储蓄还是消费。这使得政策制定者可以测试更多不同的政策方案，以实现预期的经济目标。

该项目由南洋理工大学的四位研究人员开发，是全球 14 个获得 OpenAI 资助的项目之一，这是该公司致力于通过人工智能促进经济机会和社会韧性的承诺的一部分。

这 14 个项目是从 OpenAI 在 4 月份发起的提案征集活动中收到的 400 多个提案中选出的。

由南洋商学院教务长信息系统讲席教授权赫九 (Hyeokkoo Eric Kwon) 领导的南洋理工大学团队将获得 10 万美元（126,800 新元）。

他表示，这笔资金将用于在未来六个月内开发和测试该工具（尚未命名）。

南洋理工大学团队的人工智能预测系统利用来自韩国一款移动预算应用程序超过一百万用户的真实交易数据进行训练，权教授拒绝透露该应用程序的名称。

数据包含有关用户在 2023 年至 2025 年间的消费金额、消费零售商以及每月收入的信息。

该应用通过追踪银行在用户进行交易时发送的通知、短信和电子邮件来获取数据。姓名和银行账号已被移除。消费金额、收入和购买时间也被归类到更大的范围内。

南洋理工大学团队的人工智能工具只接收匿名化数据，以防止识别用户身份。

“绝不应该出现数据只能与一个用户画像匹配的情况。根据数据的敏感程度，如果至少有三、五或十个用户拥有完全相同的特征，那么该用户画像就足够匿名化了。”权说道。

然后利用人工智能代理来模拟家庭对政府政策提案可能作出的反应。

权表示，例如，如果政府计划向每位公民发放 100 美元的现金券，该系统旨在估算此举可能产生的额外支出、这些支出将用于何处以及如果采取不同的政策，结果会如何变化。

他表示，由于人工智能能够理解复杂的模式，并且对人类行为有着广泛的了解，因此它能够比基于确定性规则的传统模拟更好地预测人类行为。

“通过定性访谈和调查来制定这些规则，但人们的实际行为与他们在调查中的回答之间往往存在差距，”Kwon说道，并补充说，收入和年龄等因素之间的复杂关系无法在传统的模拟模型中得到体现。

为了确保该系统不会产生幻觉，并且能够准确预测现实世界的结果，它将在 2025 年根据实际支出数据进行验证。

这些数据是在韩国政府于 2025 年 7 月和 9 月发放两批现金补贴前后记录的。这些补贴必须在 2025 年 11 月之前在指定的零售商处使用。

NTU 团队获得的 10 万美元资助中，一半将以现金形式发放，另一半将以 OpenAI 人工智能模型订阅计划的积分形式发放。

到 2027 年 2 月，南洋理工大学的研究人员的目标是使该工具适用于任何人群。

该工具将开源，它还能分析兴趣和政治观点等情绪，以提高预测准确性，并解释其得出结论的过程。

人工智能

技术与研究