Decision trees at work

For investors, one of the biggest challenges in stock selection is separating the signal from the noise. Individual company characteristics such as valuation, profitability, financing or analyst sentiment can all influence returns, but their impact often depends on the context in which they appear.

Understanding those interactions is a central part of MDT's investment process. For the past 25 years, we have used decision trees—a form of machine learning—to analyze historical market data and identify combinations of characteristics that have been associated with stronger stock performance.

Rather than relying on a single factor or a fixed formula, decision trees adapt their analysis to different types of companies, helping our model uncover relationships that may not be obvious through traditional approaches.

Inside the machine

Decision trees ask a structured sequence of simple yes‑or‑no questions. With MDT’s process, the goal of those questions is straightforward: to separate companies into two groups – those that have historically outperformed and those that historically underperformed. 

However, rather than asking the same questions in the same order for every type of stock, the model modifies its line of questioning based on what it learns about each type of company along the way. 

Think of it as a self-adapting line of questioning: At each step in the tree, the algorithm evaluates every available factor and a wide range of thresholds to determine which single question most improves its ability to separate outperformers from underperformers – and it does this within the framework of what it already has learned from earlier questions. 

The chosen factor used at every step is not just statistically strong in isolation; it’s the best next question to ask for that specific segment of remaining companies, because it meaningfully refines the expected return outlook from that point forward. 

A practical example

Take the case of a hypothetical firm, Company ABC. Here, the decision tree’s first question might focus on external financing, or how much the company relies on outside debt or capital.  

Historically, companies that rely heavily on outside capital tend to behave differently than those that don’t. So, this question simply divides the investment universe into two broad groups. It’s not a pass or fail decision, and the answer does not exclude the company from consideration. It’s simply one question in a long sequence of questions that helps our team better understand Company ABC and whether or not it might offer potential for alpha. 

If Company ABC does rely on external financing, then it moves down that branch of the decision tree. In practice, this often places the company into a more growth-oriented category rather than a value-oriented one.  

From there, the tree layers in additional questions to understand how multiple factors interact.  

It might, for instance, ask if the company has avoided severe underperformance. If so, it might ask whether analysts are positive and are raising their forecasts. In this hypothetical example, historical data shows that companies exhibiting this particular combination of characteristics—reliance on external financing, stable recent performance, and very high analyst conviction—have outperformed in the near term, adding to the team’s confidence that Company ABC could do the same. 

Decision trees can capture complex interactions among investment factors that have strong economic or financial rationale but are difficult for human analysts (or traditional linear models) to identify consistently. They are well suited for this because they can ask different questions for different types of companies, rather than forcing every stock into the same fixed equation.   

While there may be more approaches to picking stocks than there are stocks, and the number of techniques that are available seems to grow by the day, we think that our time-tested, decision tree approach has many characteristics — objectivity, repeatability, flexibility, transparency — that make it a desirable tool for building portfolios that can generate excess returns across many different market environments.