We were pleased to host Denis Lapitski, Sander Aarts, and Harrison Hoffman from Trexquant Investment LP at our Financial Engineering Practitioner Seminar on Monday, October 6, 2025. Their talk, "Mixture of Expert Models in Quantitative Trading," explored how Mixture-of-Experts (MoE) methods can be applied to portfolio construction and quantitative trading. Before the main presentation, Harrison offered an overview of the firm’s focus and vision in quantitative trading, setting the stage for the technical discussion.
Denis framed portfolio construction as a pipeline of Data, Alpha, and Strategy. On data, he noted technical, fundamental, news, sentiment, analyst, macro, and alternative sources, all converted to daily values (one value per stock per day). On alpha, he described it as a compact expression of data that produces a market-neutral portfolio, and explained that combining many alphas creates a tradable strategy. He also broke down strategy into selection, combination, and optimization; combinations can be linear or nonlinear. He used an instructive toy example to visualize how MoE works, introduced MoE models (from basic linear models to MoE neural networks for portfolio construction), and outlined the notion of netting and how it works.
Building on this, Sander showed how MoE—including MoE neural networks—supports portfolio construction with practical risk controls. He detailed regularization; the formation of experts (and how they aid regularization and hedging); netting as a risk-management tool; and the mechanics and benefits of forming experts by bucketing/grouping alphas to improve stability. He closed with balanced takeaways on the benefits and trade-offs of MoE in real-world workflows.
On behalf of the IEOR Department at Columbia Engineering and the entire Columbia University community, we extend our sincere thanks to Denis, Sander, and Harrison; to Jiaqi Li and Christine Chan for coordinating the seminar; and to Zadia Rutty-Turner and Afsana Rahman for their assistance with the event.
