JOIM Conference Singapore

Open Trading (OT) on the largest electronic corporate-bond request-for-quote (RFQ) platform allows investor-to-investor trading. Despite only 2% of trades occurring directly between investors, OT succeeds in attracting new liquidity providers (7%) and creating new client-dealer links (3%). These result in OT improving prices by 3bps on average, 1bp from OT winning, and 2bps from dealers improving bids, which reduces trading costs by 15-30%. However, OT worsens prices in RFQs with few dealers, which have the largest winner’s curse. A model and counterfactual experiments suggest how market redesign could mitigate adverse selection and improve liquidity in corporate bond markets.

Speaker:
Terry Hendershott, UC Berkeley Haas School of Business, Professor, Willis H. Booth Chair in Banking and Finance

Details of the discussant will be shared soon.

Moderator:
Prakash Kannan, GIC, Chief Economist and Director Total Portfolio Strategy

Panellists:
Scott Richardson, Acadian Asset Management, Senior VP and Director Systematic Credit; London Business School, Professor

Steve Brice, Standard Chartered, Global Chief Investment Officer

Tan Sin Mui, NTU, Chief Investment Officer

Speaker:
Robert Engle, NYU Stern School of Business, 2003 Nobel Laureate (Economics), Professor Emeritus of Finance

Speaker:
Ananth Madhavan, UC Berkeley Haas School of Business, Executive Director

Details of the discussant will be shared soon.

Foundation models are metered and sold in tokens — a unit that is to AI what the kilowatt-hour is to electricity. I provide a framework in which users demand tasks while providers post a low-dimensional menu of token prices, showing why a model that is cheaper per token is often more expensive per task, and how providers should set markups across input, output, cache, and reasoning tokens. The dynamics matter most for investors: because the same tokens used in production are also burned prototyping new applications, falling token prices expand the set of AI applications that exist — and when the opportunity frontier is thick enough, total token demand and revenue rise faster than prices fall, a modern Jevons regime with direct implications for compute and data-centre demand. Under competition, workload portability determines where rivalry compresses margins and where pricing power survives: portable workloads are commoditised (increasingly by AI routers), while ecosystem-bound workloads sustain both markups and heavily subsidised launch pricing. The paper closes with practical guidance for enterprise AI procurement, provider pricing, and the evaluation of AI business models.

Speaker:
Will Cong, Nanyang Technological University, President’s Chair Professor of Finance, Computing and Data Science, Director of Global Institute of Finance, Technology and Society (GIFTS)

Discussant:
Yiping Lin, Alt Data Tech, CEO

Moderator:
Henry Shen, BlackRock, Senior Investment Strategist of Systematic Active Equity Division

Panellists:
Alan Lim, MAS, Head of Financial Infrastructure and Artificial Intelligence Office

Kevin Chang, Temasek, Managing Director and Head of Quantitative Strategy and Performance Analytics

Successfully solving an applied Reinforcement Learning (RL) problem requires three distinct forms of effort: understanding the problem domain, designing the Markov decision process (MDP), and implementing a code repository that handles reward shaping, hyperparameter selection, environment validation, training, and evaluation. Recent agentic systems with sophisticated tool-use harnesses have shown that long-horizon implementation tasks can be completed with increasing autonomy in other domains. We find, however, that RL exposes a sharper bottleneck: autonomous code generation addresses only the implementation stage, while failures often originate earlier, when human intent has not yet been translated into a precise problem definition. Chat Assisted Reinforcement Learning (CARL) is designed for this specification gap. CARL guides a domain expert through an interactive seven-step workflow that formalises the problem before autonomous implementation begins, using bespoke skill files to structure the discussion for each step and compact summaries to carry decisions forward. After the formal problem description is agreed upon, CARL’s tool-use loop generates, executes, debugs, and iterates on a runnable RL training pipeline. We describe the CARL architecture, its skill-driven workflow, its boundary between interactive specification and autonomous coding, and the current limitations and future research directions for human-agent co-design of RL systems. Using realistic examples such as dynamic portfolio optimisation, we also show the drawbacks of using fully autonomous RL code generation without first resolving the specification gap.

Speaker:
Sanjiv Das, Santa Clara University, William and Janice Terry Professor of Finance and Data Science

Discussant:
Aleksander Sobczyk, Two Sigma, Senior VP