MDT's Transparent Decision Trees Drive All-Weather Quantitative Returns

Deep Dive

Daniel Mahr's early path into investing began with a fascination for numbers, evident from his childhood habit of poring over numerical data in newspapers. This interest evolved during his college years at Harvard in 1998, where he identified an opportunity in the dot-com bubble by flipping IPOs. He managed to secure allocations to these rapidly appreciating stocks, initially viewing the small share amounts as significant gains. This experience, however, also led to substantial losses when he deviated from his original investment theses, convincing him of the need for a more disciplined and systematic approach, which he found in quantitative investing.

The discussion then shifts to Mahr's entry into the quantitative investing space. He joined MDT Advisors in 2002 as a junior analyst, a firm that had been a pioneer in quant investing since 1991. After MDT was acquired by Federated Investors (now Federated Hermes) in 2006, Mahr took over leadership of the team in 2008. He has since guided the group through significant changes in data availability, computing power, and investment methodologies over two decades.

Mahr elaborates on the evolution of quantitative investing, noting that while some early strategies persist, the field has seen an explosion in processing power, data, and algorithms. This has led to a dramatic increase in the sophistication and variety of strategies that can be implemented, supported by these technological advancements.

He then details MDT's investment strategies, explaining that in 2002, their approach involved traditional factor tilting strategies using a limited number of characteristics. However, these strategies experienced difficulties during the 1998-1999 period when value and quality factors were not favored by the market. This led the firm to seek a more differentiated approach, resulting in the development of their decision tree framework, which they continue to refine today.

The concept of a decision tree approach to stock selection is explained by analogy to life insurance, where a series of yes/no questions about characteristics predict an outcome like longevity. In the stock market context, these questions pertain to company characteristics, with subsequent questions evolving based on relevance to specific company types. Mahr emphasizes that MDT's approach is not a "black box" but rather a "glass box," allowing for transparency into how the models operate and drive decisions.

Mahr articulates MDT's investment philosophy, which centers on the belief that a disciplined quantitative approach can yield an analytical advantage, leading to superior, all-weather portfolio returns. He contrasts this with a "global macro crystal ball" approach, stating that MDT's method relies on diversification across companies with differentiated alpha drivers rather than predicting market environments.

Regarding differentiation in a crowded market, Mahr highlights MDT's long-standing use of machine learning and artificial intelligence since 2001, giving them a significant head start. He notes that while many competitors are now exploring these technologies, MDT has accumulated extensive knowledge regarding their advantages and potential pitfalls in noisy data environments like stock return forecasting.

The critical challenges of overfitting and underfitting data in model building are discussed. Mahr explains that while a simple model might avoid overfitting, it risks underfitting and leaving explanatory power on the table. Striking the right balance between model complexity and simplicity is a key focus that MDT has evolved over decades.

Mahr describes how research ideas are generated, drawing from academic and practitioner literature, but more frequently from observations of their own strategies' behavior over their 30-plus years of investing. These observations across various market cycles have informed significant process enhancements.

He introduces "company age" as an unusual factor used in their models, measuring how long a company has been publicly traded. While company age alone does not predict performance, Mahr explains that the decision tree framework leverages it to ask different questions of young companies versus established ones, with valuation being more critical for older companies. The selection of factors is driven by the investment team's research, but their application within the algorithm is mechanically determined.

The process for retesting and potentially removing factors is detailed. Factors are removed if they no longer perform or if new factors capture a correlated effect. The decision to remove a factor is data-driven, stemming from observations of fewer trades being driven by that factor over time. Research projects then confirm whether removing the factor impacts portfolio outcomes negatively.

Mahr visualizes the internal workings of the "glass box" as a "forest of trees" rather than a single tree, a progression enabled by increased processing power and advanced algorithms. He explains that in the past, a single tree was printed and taped to the wall for trade review, but with thousands of trees, analytical tools are used to synthesize and summarize their operations. Each tree typically involves two to five decision points, as deeper trees operate on progressively smaller data pools, limiting their depth.

An example of a decision tree path is provided, starting with a question about a company's financing activities (issuing debt or shares versus buying them back). Companies with high financing levels tend to underperform, but the algorithm seeks out strong momentum companies within this group that can outperform despite financing. Down the non-financing branch, other differentiated company groups are identified. Subsequent questions might involve volatility or company age, with momentum being more meaningful for newer companies. The stopping rules for asking questions are mechanical, with a hard limit of five questions per tree or stopping if a branch's data pool becomes too small.

The aggregation of thousands of tree signals into a portfolio is managed by a portfolio optimizer. This optimizer considers alpha forecasts from the decision tree model, risk management through hard constraints and a statistical risk model, and trading costs, aiming for portfolios with consistent outcomes and efficient trading.

Mahr discusses the trade-off between alpha signals and market impact, noting that less liquid companies will have smaller positions and slower trading speeds. He also addresses human judgment in overriding model decisions, emphasizing a data-oriented approach. When a company reports earnings, the team analyzes how those earnings will impact the model's factors and subsequent tree decisions, aiming for precision in potential overrides.

The reflexivity of other quantitative participants is considered in the context of leverage. Mahr points to historical events like Long-Term Capital Management and the 2007 quant quake, attributing them to periods of underperformance magnified by leverage, rather than inherent flaws in quantitative strategies themselves. He notes that statistical arbitrage strategies had become crowded, leading some managers to increase leverage.

Mahr observes changes in market structure, suggesting that markets may have shifted from a continuous trend toward efficiency to something different in the last couple of years. He speculates that this could be due to the rise of "pod shops," passive management, retail trading, or a combination of factors, but stresses MDT's focus on having active strategies that can capitalize on inefficiencies regardless of their source.

A typical day involves downloading updated data, recalculating characteristics, running companies through the forest of trees for updated forecasts, and re-optimizing portfolios to generate a trade list. The day begins with a trade review process focused on verifying data accuracy, understanding model dynamics, and identifying any market news not yet captured by the data inputs. The rest of the day is dedicated to research, idea generation, and model improvement, utilizing proprietary tooling built in-house since 1991.