Events

Mimicking Finance

Abstract:

We use advances in AI and machine learning to identify which behaviors of key economic agents are predictable from their past actions. Although these decisions may appear novel, we find that 71% of mutual fund managers’ trade directions can be predicted before they make a trade, and for some managers, nearly all trades in a quarter are predictable. Predictability is higher for managers with longer trading histories and in less competitive categories, and lower for managers with larger personal stakes in their funds. Importantly, less predictable managers and positions significantly outperform more predictable ones. Our framework separates predictable behavior from genuinely novel decisions, allowing researchers to evaluate where financial agents create or destroy value.

Bio:

Lauren Cohen is the L.E. Simmons Professor in the Finance & Entrepreneurial Management Units at Harvard Business School and a Research Associate at the National Bureau of Economic Research. He is an Editor of the Review of Financial Studies, along with being a past Editor of Management Science, and serving on the editorial board of the Review of Asset Pricing Studies.

Professor Cohen teaches in the MBA Program, Executive Education Program, Doctoral Program, and Special Custom Programs at the Harvard Business School, teaching across Family Enterprise, Investment Management, and Innovation Course Offerings. In particular, he is the Faculty Co-Chair and Designer of the HBS Executive Education course ‘Building a Legacy: Family Office Wealth Management,’ designer of a first-of-its-kind MBA Course in Family Offices entitled ‘How to Not Bankrupt Your Family,’ and Faculty Co-Chair and Designer of the HarvardX Fintech course.

He is an award-winning researcher, and best-selling case writer, with works published in the top journals in Finance and Economics. His work is frequently profiled in various media outlets including The Wall Street Journal, The New York Times, The Washington Post, The Economist, and Forbes. It has also been recognized by numerous National Science Foundation (NSF) Awards, including a National Science Foundation Early Career Development (CAREER) Award for his research agenda on Relationships in Finance. He was named a 2008 Pensions & Investments “Cutting Edge Academic,” a Top 40 Under 40 Business School Professor in 2017 by Poets & Quants, and a top teacher at Harvard by CNBC.
 

Limits to (Machine) Learning

Abstract:

Machine learning (ML) methods are highly flexible, but their ability to approximate the true data-generating process is fundamentally constrained by finite samples. We characterize a universal lower bound, the Limits-to-Learning Gap (LLG), quantifying the unavoidable discrepancy between a model's empirical fit and the population benchmark. Recovering the true population R 2 , therefore, requires correcting observed predictive performance by this bound. Using a broad set of variables, including excess returns, yields, credit spreads, and valuation ratios, we find that the implied LLGs are large. This indicates that standard ML approaches can substantially understate true predictability in financial data. We also derive LLG-based refinements to the classic Hansen and Jagannathan (1991) bounds, analyze implications for parameter learning in general-equilibrium settings, and show that the LLG provides a natural mechanism for generating excess volatility.

Bio:

Semyon Malamud is Swiss Finance Institute Associate Professor at EPFL and Research Fellow Center for Economic Policy Research. He is also Associate Editor of the Journal of Finance.

Discovering Multiagent Learning Algorithms with Large Language Models

Abstract:

Much of the advancement in Multi-Agent Reinforcement Learning (MARL) for imperfect-information games has historically depended on the manual, iterative refinement of algorithmic baselines. Recently, evolutionary coding agents powered by Large Language Models (LLMs) have emerged as powerful tools to automate this discovery process. In this work, we deploy one of such agentic frameworks, AlphaEvolve, to navigate the design spaces of two distinct game-theoretic paradigms: counterfactual regret minimization (CFR) and policy-space response oracles (PSRO). This automated search yielded two algorithms: Volatility-Adaptive Discounted (VAD-) CFR and Smoothed Hybrid Optimistic Regret (SHOR-) PSRO, which are consistently competitive with state-of-the-art human-designed baselines across an 18-game evaluation suite spanning Poker, Goofspiel, Liar's Dice, Blotto, and Battleship variants. However, because the LLM optimizes for fitness on a specific training set, it often constructs highly synergistic, complex mechanisms tailored to those environments. Through systematic ablation studies, we demonstrate that while these mechanisms are tightly coupled, the true driver of generalization lies in a minimal algorithmic core. By distilling the LLM's discoveries down to their most fundamental principles, we produce two minimal solvers: Warm-started Optimistic Predictive (WOP-)CFR and Projection Matching (PM-)PSRO. These distilled versions achieve superior performance on generalization with greatly reduced structural complexity, providing a clear methodology for using LLMs in algorithmic discovery.

Bio:

Zun is a Research Engineer at Google DeepMind. He earned his Ph.D. in Computer Science from the University of Michigan in January 2024, where he was advised by Prof. Michael P. Wellman.

His research focuses on studying intelligence through a multi-agent lens. He has worked on computational game theory, deep multi-agent planning and learning, agent-based simulation methods, auctions, and data markets.

Posters will be available after the conference.

The Uncertainty of Machine Learning Predictions in Asset Pricing

Abstract:

Machine learning in asset pricing typically predicts expected returns as point estimates, ignoring uncertainty. We develop new methods to construct forecast confidence intervals for expected returns obtained from neural networks. We show that neural network forecasts of expected returns share the same asymptotic distribution as classic nonparametric methods, enabling a closed-form expression for their standard errors. 

We also propose a computationally feasible bootstrap to obtain the asymptotic distribution. We incorporate these forecast confidence intervals into an uncertainty-averse investment framework. This provides an economic rationale for shrinkage implementations of portfolio selection. Empirically, our methods improve out-of-sample performance.

Bio:

Linda Schilling joined Olin Business School as an Assistant Professor in Finance in 2021. Linda received her PhD in Quantitative Economics from the University of Bonn in 2017. Since then, she has been holding positions as Assistant Professor for Financial Economics at the University of Utrecht in the Netherlands and Ecole Polytechnique CREST in Palaiseau, France.

Linda holds a Diplom (equiv. to Masters) in pure mathematics from the University of Bonn and a Master's in financial mathematics from the University of Edinburgh.

Learning Beyond the Training Distribution 

Abstract:

Transformers have illustrated remarkable capability to learn new concepts from context at inference time. In settings undergoing distribution shift, this is particularly important since it allows the model to use the current context to condition its behavior under the new distribution. However, the interplay between the pretraining distribution and the ability of the model to extrapolate to new settings is poorly understood. Understanding how to properly design a pretraining distribution allows for more efficient model development, especially in the case of numerical tasks with limited data. We first study conditions that allow general deep learning models to extrapolate outside their training domain using techniques from extreme value theory. We then discuss how the shape of the pretraining distribution controls how well the model learns under distribution shift given new context information. Finally, we illustrate these results on a number of tasks involving strong distribution shift including stochastic processes and Monte Carlo integration. The theory predicts how to choose the pretraining distribution to maximize performance over a wide range of conditions, with particular emphasis on numerical tasks. 

Bio:

Ali Hasan is a machine learning researcher at Morgan Stanley and a research affiliate at Duke University.

His research focuses on bridging the gap between the computational aspects of deep learning and the theoretical foundations of applied mathematics, including stochastic processes, differential equations, and extreme value theory. He is interested in applying these methods to model phenomena in biomedical, environmental, and financial applications.

Machine Learning Meets Markowitz

Abstract:

The standard approach to portfolio selection involves two stages: forecast the asset returns and then plug them into an optimizer. We argue that this separation is deeply problematic. The first stage treats cross-sectional prediction errors as equally important across all securities. However, given that final portfolios might differ given distinct risk preferences and investment restrictions, the standard approach fails to recognize that the investor is not just concerned with the average forecast error - but the precision of the forecasts for the specific assets that are most important for their portfolio. Hence, it is crucial to integrate the two stages. 

We propose a novel implementation utilizing machine learning tools that unifies the expected return generation process and the final optimized portfolio. Our empirical example provides convincing evidence that our end-to-end method outperforms the traditional two-stage approach. In our framework, each investor has their own, endogenously determined, efficient frontier that depends on risk preferences, investor-specific constraints, as well as exposure to market frictions.

Bio: 

Campbell R. Harvey is Distinguished Professor of Finance at the Fuqua School of Business, Duke University in Durham, North Carolina as well as a Research Associate of the National Bureau of Economic Research in Cambridge, Massachusetts. He served as Editor of the Journal of Finance for six years and as President of the American Finance Association in 2016.

Harvey received, in 2015, 2016, 2020, 2022 and 2023, a Bernstein Fabozzi/Jacobs Levy Award for Best Article appearing in The Journal of Portfolio Management. He has also received 10 Graham and Dodd Awards/Scrolls for excellence in financial writing from the CFA Institute. He has published over 170 scholarly articles as well as four books on topics spanning active investing, risk management, emerging markets, behavioral finance, financial econometrics and computer science. His recent books include Strategic Risk Management as well as DeFi and the Future of Finance (both John Wiley and Sons). Professor Harvey teaches both an advanced asset management course, as well as an offering that focuses on decentralized technologies and AI.

Harvey is a Founding Director of the Duke-CFO Survey. This influential quarterly survey is now co-administered by the Federal Reserve Bank.

Harvey serves as Partner and Director of Research at Research Affiliates, LLC who oversees over $150 billion in client funds as well as Investment Strategy Advisor to the Man Group plc, the world largest, publicly listed, global hedge fund provider. He served as Digital Asset Advisor to BlackRock in 2022. 

He has served on the faculties of the Stockholm School of Economics, the Helsinki School of Economics, and the Booth School of Business at the University of Chicago. He has also been a visiting scholar at the Board of Governors of the Federal Reserve System. He holds a Ph.D. from the University of Chicago.

Over the past ten years, Professor Harvey has taught Innovation and Cryptoventures at Duke University. The course focuses on blockchain technology and decentralized finance. He also teaches an advanced asset management course. He also offers four-course DeFi specialization on Coursera with a total enrollment of 102,000.

Learning to Discover at Test Time

Abstract:

How can we use AI to discover a new state of the art for a scientific problem? Prior work in test-time scaling, such as AlphaEvolve, performs search by prompting a frozen LLM. We perform reinforcement learning at test time, so the LLM can continue to train, but now with experience specific to the test problem. This form of continual learning is quite special, because its goal is to produce one great solution rather than many good ones on average, and to solve this very problem rather than generalize to other problems. Therefore, our learning objective and search subroutine are designed to prioritize the most promising solutions. We call this method Test-Time Training to Discover (TTT-Discover). Following prior work, we focus on problems with continuous rewards. We report results for every problem we attempted, across mathematics, GPU kernel engineering, algorithm design, and biology. TTT-Discover sets the new state of the art in almost all of them: (i) Erdős’ minimum overlap problem and an autocorrelation inequality; (ii) a GPUMode kernel competition (up to 2× faster than prior art); (iii) past AtCoder algorithm competitions; and (iv) denoising problem in single-cell analysis. Our solutions are reviewed by experts or the organizers. All our results are achieved with an open model, OpenAI gpt-oss-120b, and can be reproduced with our publicly available code, in contrast to previous best results that required closed frontier models. Our test-time training runs are performed using Tinker, an API by Thinking Machines, with a cost of only a few hundred dollars per problem.

Bio:

Federico Bianchi is a Staff ML Scientist at TogetherAI, working on self-improving agents. He was previously an AI Engineer at OpenEvidence and a post-doc at Stanford University. His work has been published in major journals such as Nature and Nature Medicine and conferences such as ICLR, ICML and ACL.