What You Need to Know

With mounting risks in financial markets, cluster analysis should play an increasing role in an investing toolkit. Here’s what investors need to know about how it works—and what to ask asset managers—to know whether it’s being applied effectively to an equity portfolio.

From anthropology to politics, analysts in many fields have used cluster analysis to help decipher complex relationships for nearly 90 years. But investment firms are only beginning to discover the powerful applications for detecting unknown risks lurking in market behavior patterns.

Mounting uncertainty in financial markets and increasingly knotty trading anomalies of securities are fueling demand for more sophisticated risk-management techniques. Cluster risk analysis is a good supplement to traditional risk tools and should play an increasing role in an investing tool kit, in our view. This technique seeks correlated sources of risk that may not be obvious to quantitative risk models or fundamental analysts. But investors need a better understanding about how cluster analysis works to ask their asset managers the right questions about how it’s being applied to an equity portfolio.

Why Should Investors Care About Clusters?

Cluster analysis dates to 1932, when it was first applied to an anthropological study that measured similarities between cultures. Since then, it’s been used in a long list of disciplines. In psychology, it was famously applied by Raymond Cattell to group personality traits into clusters in 1943. Biologists have used it since the 1960s to find common groups of cells and organisms. Political campaigns, market surveys and medical research all use cluster analysis to help analysts discover clear categories and to explain underlying processes and patterns.

In the investment world, cluster analysis is a relative newcomer. Recent studies from sell-side analysts have used it, for example, to generate insight about factor risks and to detect stages of the equity market cycle. Buy-side asset managers haven’t widely adopted cluster analysis in their risk-management tool kits, partly because computing technology wasn’t powerful enough to apply it to the complexity of markets. Yet as machine-learning techniques gain traction in the financial industry, more firms are starting to see the benefits of cluster analysis.

We think cluster analysis can generate essential investing insight. In today’s markets, it’s becoming increasingly difficult to predict how a plethora of political and policy risks could affect stock returns. With cluster analysis, portfolio managers can uncover relationships that other risk models may miss.

Moving Beyond Standard Risk Models

Standard risk models for equity portfolios focus on factors that affect performance, such as style, sector, country and currency. These models are designed to make sure that an investor is aware of a portfolio’s exposures and can avoid having the portfolio too heavily tied to particular areas of the market.

But fundamental risk models track only a defined set of risks. There are countless hazards to investment portfolios that slip under their radar screens, from interest rate risk to trade wars. When performance patterns of a group of stocks with similar business profiles but different risk classifications become correlated, cluster risk is created.

Cluster risk analysis aims to detect these unknown risks. It can help investors understand portfolio exposures and can be used as a building block for portfolio construction. When applied comprehensively, it can also be used as a framework for reviewing data from other risk tools to help understand the changing drivers of stock returns.

In portfolios that are driven by bottom-up stock selection, the quantitative nature of cluster risk analysis provides an important perspective. It serves as a lens through which the portfolio managers can balance the portfolio-level exposure to macroeconomic and other factors.

How Does It Work?

Cluster analysis has developed into a sophisticated machine-learning technique that segments stocks into groups whose returns have been moving closely together over a defined period. For example, it can help separate groups of stocks within an industry or subindustry that will benefit in a risk-on trade, when markets reward riskier assets, from other groups of stocks that may be more aligned with a risk-off environment. The analysis requires sophisticated data mining based on algorithms with internal rules, rather than learning from examples. The goal is to create groups of items that are “close to” and “distant” from each other.

Choosing the right algorithm for the job is no easy task. The literature on cluster analysis includes thousands of different algorithms that can be used. Finding the right one for a particular problem requires expertise in the vast mathematical and computational options as well as in the target market and the investing research objectives.

Even after getting the math right, defining clusters can be tricky. For example, as the display below shows, the same group of 15 dots can be seen as two or three distinct clusters. Now, imagine the data-mining complexity involved in identifying distinctive clusters of stock patterns among thousands of global securities that are being traded constantly, with prices shifting by the millisecond.

There are different ways to classify distinctive clusters. In partitional clustering, all clusters are evaluated at once and separated into distinctive groups. However, this approach is not always useful when clustering enormous data sets such as stock markets.

Hierarchal analysis is often the preferred method. This approach recursively finds successive clusters, using previously established clusters. It’s especially useful when flexibility is needed in choosing how many clusters you want. In the display below, cluster C3 is not similar to clusters C1 or C2. At the same time, clusters C3 and C4 show similarity to cluster C5. Longer horizontal “branches” of the tree connecting different clusters suggest less similarity between the linked clusters.

Identifying Clusters in Markets

For investment managers, detecting clusters is only the first step. Cluster analysis is a hypothesis-generating analysis, rather than a hypothesis-testing technique. In other words, it aims to create questions that require further examination rather than to resolve an existing query.

So, after discovering stock clusters with similar behavior, the first question you need to ask is, what is driving the performance patterns. Even the most advanced machine-learning technologies can’t answer this question. For this crucial stage of cluster risk analysis, experienced and skillful human interpretation are essential. Analysts can use the following frames of reference to help interpret what affects cluster behavior:

  • Exposure to factors or group memberships (e.g., regions and industries)—standard equity factors, such as valuation, profitability, growth, balance-sheet quality and size, may help explain cluster performance. But we also need to consider influences such as regional exposures, geographic revenue exposures and changes to earnings forecasts.
  • Stock-price sensitivities—such as changes in interest rates, inflation, the oil price and currency movements.
  • Cluster performance—the returns and return patterns (such as upside/downside capture) of the cluster itself can provide insight into what is differentiating about a particular group of stocks.
  • Exogenous influences—such as developments related to policy decisions or political risk, including Brexit or the US-China trade war. News events—such as vaping health concerns or the regulatory crackdown on mega-cap technology companies—will affect some stocks much more than others.

Cluster Analysis in Action

In recent years, we at AllianceBernstein have increasingly used cluster analysis to supplement our standard risk-management tools in some equity portfolios. The following examples illustrate some important insights about curious market patterns that the cluster analysis efforts of our investment management teams have revealed.

Rise of the FAANGs: As Facebook, Amazon, Apple, Netflix and Google rose to dominance in their industries in recent years, investors have flocked to the so-called FAANG stocks. But although technology drives the businesses of all five companies, they aren’t all classified by technology stocks in standard risk models. Standard risk models might check that a portfolio doesn’t have too much exposure to the technology sector or the consumer sector. But they don’t look at the FANG or FAANG groups as a whole. And if the performance of these stocks is correlated, a portfolio that holds too much—or too little—of the group might face unwanted risks.

Cluster research that we conducted in 2017 indicated that the FAANG group’s stock returns had become increasingly correlated at the time. A risk model that isn’t aware of this correlation may prompt a portfolio manager to purchase all five FAANG stocks—a position that could be vulnerable to a sharp downturn of the group. Or, the risk model may suggest adding to positions in Netflix or Amazon.com, without considering the possibility that these positions may suffer if Apple has a bad day.

From regional clusters to beta clusters: Through most of 2018, global developed markets were generally forming their own subsets of regional clusters (Display, left). Stocks in Europe, Japan and the US were all clustered together as distinctive groups.

However, as 2019 progressed, these regional groupings broke apart and instead lower-beta defensives and higher-beta cyclicals from different regions became more closely related. We were now seeing a clear high/low beta split of return patterns rather than a regional split (Display, right). At the same time, bond proxies, such as utilities, some consumer staples and real estate investment trusts, were also forming a strong and separate group of clusters. These changes suggested increased focus on diversifying beta and interest-rate sensitivity risks.

Redefining risk-on versus risk-off stocks: Cluster analysis findings can defy conventional wisdom. For example, while industrials are generally seen as risk-on stocks, some companies in the sector have business models that make them more defensive. In contrast, healthcare is often seen as a defensive sector, but some drugmakers, with a small number of products or imminent patent expiries, may be much less defensive than perceived.

In another case last year, we found a curious change in the trading patterns of a UK water utility. In the past, this stock behaved in line with its utility peers, which are typically considered defensive. At some point, it started trading in line with UK retailers, which are much more cyclical. A possible explanation? The shift may have been related to the rise of the Labour Party in UK polls, as the party has a stated policy of nationalizing assets like water utilities. This potential disruption could greatly widen the range of possible outcomes for the stock. Whatever the reason, the utility’s risk profile had changed significantly, which should prompt a rethink of the stock’s role as a defensive ballast for a portfolio. And a standard risk model probably wouldn’t have detected the change in the stock’s trading pattern because global utilities continued to behave like defensive stocks.

Tariff risks reshape Chinese stock patterns: In 2017, cluster analysis results showed us that Chinese stocks mainly moved as a group of clusters, quite separate from other regions. However, through 2018, as tariffs were announced, we saw a shift. A group of Chinese stocks, which contained consumer discretionary, internet marketplaces and technology, started moving more closely with stocks of companies that faced risks from tariffs, such as US energy companies and several global metal and mining companies.

From Cluster Analysis to Portfolio Construction

When a cluster analysis reveals themes like those described above, what should a portfolio manager do? The first step is to check whether the portfolio’s holdings in the cluster that has been identified is very different from the benchmark’s weight in the cluster.

Then, ask whether the cluster overweight or underweight is intentional. If not, the portfolio manager must assess whether the position is justified or whether an adjustment is warranted. Detecting a cluster risk may also prompt a portfolio manager to reposition in order to change factor exposures or to look in underrepresented clusters for new ideas.

Making the right changes in response to a cluster risk can provide an extra layer of downside protection, while also helping to secure the alpha potential from bottom-up fundamental research. There are countless hazards for investors to track in today’s equity markets—and a wide array of tools for managing common risks. When applied strategically and skillfully, cluster analysis can provide a risk-management advantage by helping investors manage risks they don’t even know may be hiding in a portfolio allocation.

Peter Chocian, Senior Quantitative Analyst and Portfolio Manager—Equities

Nelson Yu, Head of Quantitative Research—Equities

The views expressed herein do not constitute research, investment advice or trade recommendations and do not necessarily represent the views of all AB portfolio-management teams and are subject to revision over time. AllianceBernstein Limited is authorized and regulated by the Financial Conduct Authority in the United Kingdom.

Clients Only

The content you have selected is for clients only. If you are a client, please continue to log in. You will then be able to open and read this content.