HKML S8E1 - Quant ML

Event TBC. Looking for host sponsor.

Tentative programme:

3 speakers.

Speaker: Gautier Marti

Topic: Detecting Evasive Answers in Financial Q&A with NLP and Large Language Models

Summary:

Earnings calls and other financial Q&A sessions are a critical source of information for investors, yet executives often respond evasively without explicitly lying. This talk presents recent research, led by Khaled Al Nuaimi, on evasive answer detection as a standalone NLP task, distinct from sentiment or factuality. Gautier will introduce psychology-informed taxonomies of evasiveness, annotated datasets covering earnings calls and central bank press conferences, and both lightweight and large-model-based detection approaches. He will also discuss how recent benchmark efforts, including work by researchers based in Hong Kong, are shaping large-scale evaluation, and why automated detection of evasive language matters for transparency, market analysis, and downstream financial NLP applications.

Bio:

Gautier Marti is a quantitative researcher with experience in equity statistical arbitrage and credit trading strategies. He has a strong background in machine learning and its applications to alpha research. Alongside his professional work, Gautier actively follows and experiments with advanced research to evaluate its practical impact on portfolio performance. He also mentors several PhD students on topics including modern NLP and multimodal techniques for finance, explainability of trading models, and forecasting using large language models and network-based approaches.

Speaker: Yizhi Song

Topic: Deep Tangency Portfolio

Summary:

This talk introduces a deep learning framework for estimating tangency portfolio weights by augmenting benchmark factors with a long–short deep factor learned from high-dimensional firm characteristics. The deep factor is optimized to maximize Sharpe ratio, hedging benchmark risks while capturing nonlinear signals across multiple markets. Applied to U.S. corporate bonds using 132 firm-level characteristics spanning bond, equity, and option data, the resulting deep tangency portfolio achieves an out-of-sample Sharpe ratio above 2 when benchmarked to the market factor, outperforming portfolios based on common observable or latent factors. The results suggest that bond risk premia are dense and require nonlinear, multi-market integration to effectively span the efficient frontier.

Bio:

Yizhi Song is a Quantitative Researcher at Squarepoint Capital in Hong Kong, where he focuses on systematic fixed income strategies. He holds a PhD in Empirical Asset Pricing from City University of Hong Kong (2024), with research applying statistical methods and machine learning to asset pricing. Yizhi also earned an M.S. in Applied Mathematics from Northwestern University and a B.S. in Biomedical Engineering from Huazhong University of Science and Technology.