Mingwei Ma
Jako Ma
Managing Partner & Chief Investment Officer, Capibara Capital
- Email jakoma@capibaracapital.com
- GitHub mingweima
- X @ma5_mingwei
I am co-founder and CIO of Capibara Capital, a Hong Kong-based systematic trading firm, where I lead quantitative research end to end. Before that I was a quantitative researcher at Ubiquant Investment, working mainly on applying deep learning to stock, futures and convertible bond trading.
I studied Physics and Philosophy and then Mathematical and Theoretical
Physics at the University of Oxford, and was a Ph.D. student in
Statistics and Econometrics at Chicago Booth before leaving to work
full-time. My research has covered deep reinforcement learning,
multi-agent coordination and machine learning for finance. These days
I am also interested in AI entrepreneurship and the agent-centric
internet, and I build small open-source tools for AI agents, such as
brainlib.
News
- Oct 2026 Released brainlib v0.2.1, an open-source brain library for AI agents.
- May 2025 Joined Capibara Capital as Managing Partner & CIO.
- 2023 Learning Intuitive Policies using Action Features appeared at ICML 2023.
Experience
-
Capibara Capital (Hong Kong) Limited, Managing Partner & Chief Investment Officer2025 – present
-
Ubiquant Investment, Beijing, Quantitative Researcher2022 – 2024
Education
-
University of Chicago, Booth School of Business, Ph.D. studies in Statistics and Econometrics (left to work full-time)2020 – 2022
-
University of Oxford, M.Sc. in Mathematical and Theoretical Physics2019 – 2020
-
University of Oxford, B.A. in Physics and Philosophy, First Class Honours2016 – 2020
Publications
-
Learning Intuitive Policies using Action Features
M. Ma, J. Liu, S. Sokota, M. Kleiman-Weiner, J. N. Foerster
ICML 2023 -
Solving and Estimating Structural Dynamic Equilibrium Models via Multitask Reinforcement Learning
M. Ma, C. Dai, D. Xiu
Working paper, 2022
Projects
brainlib: an agent-agnostic brain library
A harness that keeps a local Markdown folder any local AI agent (e.g. Cursor, Claude Code, Codex) can read and write, so personal context and memory survive switches between models and projects. Three modules: capturer buffers preferences and decisions with their source, organizer turns the buffer into proposed edits for review, and maintainer runs periodic health checks and archiving.
Honors
- Booth Ph.D. Fellowship, University of Chicago2020 – 2022
- Chandrasekar Prize for Best Performance in Physics, University of Oxford2020
- Gibbs Prize for Best Performance in Physics, University of Oxford2019
- Sterling Scholarship for Academic Excellence, University of Oxford2018
- National First Prize, Chinese Physics Olympiad (31st and 32nd)2014, 2015