On July 25, 2026, Samuel Drapeau, Director of the Quantitative Finance Research Center (QFRC) at Shanghai Jiao Tong University, took the stage at Shanghai Tower during the AI & Global Investment Ecosystem Dialogue and the Fifth China Quantitative Investment White Paper Symposium. Guided by the Lujiazui Financial Salon Secretariat and co-organized by BigQuant, Huatai Securities, FinElite, INGRAM, and Alibaba Cloud, the event brought together parallel forums, institutional exhibitions, academic exchanges, and private-fund roadshows. Drapeau's keynote, "Human–AI Co-Research: From Artificial Intelligence to an End-to-End Simulation World," boiled down to one idea: train AI to trade in a simulated market, not in a live one.
Any quant who has built a system knows that backtest performance does not guarantee live performance. An order changes liquidity. A partial fill changes inventory. Timing changes information. And every other participant reacts. None of that can be captured in a static dataset. Drapeau made this the crux of his talk: a financial AI agent must be evaluated as a market participant, not as a prediction model facing a fixed target.
To meet that requirement, the QFRC team has been building the next version of FastBox — a controlled multi-agent reinforcement learning environment. A matching engine, data replay, and an agent runtime form the system's backbone. On top of it, different trading strategies enter the same order book as competing agents, sensing each other and reacting. Market liquidity, strategy crowding, and price impact then emerge naturally rather than being imposed by a formula. Drapeau was candid about progress: the Rust execution core, Python research layer, order lifecycle, and risk controls are in place, while multi-agent scheduling and learned market generation remain under integration.
The platform is a collaborative effort spanning Yiqing Lin, Mathieu Laurière, Paul Weng, and external collaborator Ariel Neufeld, covering platform architecture, financial statistics, uncertainty modeling, stochastic control, and mean-field games. Drapeau closed with a clear message: before financial AI touches real capital, the evidence must be reproducible and humanly auditable. "Financial AI should enter real markets with evidence — not hope."
