Daniel Mahr – Glass Box Quant at MDT Advisers (EP.472) | Capital Allocators – Inside the Institutional Investment Industry Brief

Key Takeaways

Quantitative investing strategies have significantly evolved over two decades, incorporating advanced computing and data.
MDT employs a 'glass box' decision tree framework, blending transparency with sophisticated stock selection.
The firm balances analytical rigor with human judgment, using data to inform and validate discretionary overrides.
MDT's primary focus is on generating an analytical edge, rather than pursuing informational advantages.
Market structure shifts and factors like leverage present ongoing challenges and opportunities for systematic strategies.
Artificial intelligence is explored for team efficiency and idea generation, not yet for direct stock selection.

Daniel Mahr leads MDT, Federated Hermes' $26 billion quantitative equity group.
He began as a junior analyst at MDT in 2002 and assumed team leadership six years later.
Mahr's early investing experience included flipping IPOs during the dot-com bubble as a college student.
This speculative experience underscored the importance of a disciplined, systematic investment approach, which he found in quantitative finance.

MDT's investment philosophy emphasizes a disciplined quantitative approach to stock picking.
The objective is to gain an analytical advantage and generate all-weather portfolio returns through diversification.
MDT has utilized machine learning tools since 2001, providing a significant head start over competitors.
Early adoption of machine learning led to a deep understanding of potential pitfalls, such as overfitting and underfitting data.

MDT describes its investment approach as a 'glass box,' ensuring transparency in decision-making despite sophisticated machinery.
The quantitative model unexpectedly identified powerful reversal effects in stocks that had experienced significant price drops.
The model's trading of deeply discounted stocks, initially met with skepticism due to negative company news, validated the strategy by exploiting human emotional biases.
Research ideas are sourced from academic literature and from the team's observations of strategy performance over decades of market cycles.

The investment team drives the selection of factors for the model, while the algorithm mechanically determines their usage and weighting.
Factors are removed from the model if they become ineffective due to market evolution or if a new factor captures a similar effect.
An example cited is the removal of book-to-price as a factor from the model.
Factor removal decisions are data-driven, often stemming from observing a decrease in trades attributable to that factor over time.
The process includes evaluating the impact of removing a factor on backtested returns and risk, leveraging the model's 'glass box' transparency.

MDT's decision trees evolved from a single tree to a 'forest of trees' to avoid overfitting data, with each tree typically asking two to five questions.
Initial questions often relate to company financing activities, as firms with significant financing tend to underperform.
The model identifies a subset of high-financing companies that outperform, generally the strongest momentum companies, despite financing needs.
Subsequent questions involve evaluating volatility, where consistent stock price increases are favored, and company age, with momentum being more meaningful for newer firms.
Stopping rules are mechanical, with a hard limit of five questions per tree and a condition to stop if a branch has too small a data pool.

Human judgment is integrated by evaluating how discretionary overrides align with data-driven insights, particularly when company earnings reports are released.
The 'glass box' approach allows for precise overrides based on how specific factors influence model decisions.
The conversation addresses reflexivity in quantitative investing, noting that historical events like the 2007 Quant Quake demonstrated how leverage magnified underperformance.
Market structure has evolved, and the diminishing edge from traditional factor tilting over the past decade has made systematic investing more challenging.

MDT's daily workflow begins with downloading updated data, recalculating characteristics, and running company forecasts through decision tree models.
This process leads to the daily re-optimization of portfolios and the generation of trade lists.
The daily trade review focuses on verifying data accuracy and understanding model drivers, while also accounting for potential external news not yet reflected in data inputs.
MDT distinguishes between informational and analytical edges, prioritizing analytical advantages through decision trees and machine learning.
MDT's models are trained on approximately 50 years of historical financial and price data to enhance robustness across various market cycles.

Large language models like ChatGPT are not currently used in MDT's modeling due to concerns about training data and out-of-sample testing, particularly the risk of models knowing future outcomes.
Federated Hermes is exploring AI software development co-pilots to improve team efficiencies, recognizing AI's value in enhancing productivity and idea generation within their proprietary system.
Current research efforts include revisiting stock ownership as a factor for stock selection and further exploring AI applications for enhanced productivity and idea generation.
Daniel Mahr identifies team building as a significant challenge due to the high demand for data science and AI talent, though a recent softening in the software programming market offers new hiring opportunities.
Mahr expresses continued excitement for his work, citing the dynamic nature of financial markets and constant introduction of new information.