Quantitative Science—Actively Adding To Fixed Income Decisions

Quantitative vs. active debate? This Fixed Income Markets explains how our “active quant” approach sets us apart.

The Franklin Templeton Fixed Income Group

In this Issue

Fixed income investing has undergone a sea change in the past decade. By tossing out some active management orthodoxies and embracing new technologies and quantitative techniques, we believe some managers are better equipped to capture unique insights and excess returns for their clients.

Several investment shops, however, have cast this transformation as a binary choice—a confrontation between enlightened quantitative approaches and outmoded “traditional” active managers who remain steadfast in the fundamental economic laws of capital markets. We think this quantitative vs. active debate sets up a false dichotomy. In this paper we explain how our active approach to quantitatively derived insights sets apart our “active quant” investment process.

Key takeaways

  • We start by explaining the origins of factor investing and the dawn of machine learning techniques that are upsetting the status quo in fixed income portfolio management.
  • We discuss top-down macroeconomic research and explain how we use machine learning algorithms to bring objectivity and discipline to the process of forecasting asset prices across a multi-sector fixed income universe.
  • We explore the bottom-up process of security selection—comparing factor-driven methods that capture persistent quantifiable drivers of excess returns, with traditional credit security analysis. We explain why deep credit research remains critical to the process of assessing risk premia (i.e., expected returns in excess of the risk-free rate).

Active quant fixed income

Quantitative trends in fixed income

Investing in fixed income markets has undergone a big transformation in recent years. For starters, factor investing—originally popularized by index providers like MSCI and commonly found in equity portfolios—is now spreading into fixed income.1 Sometimes called “smart beta” or “systematic investing,” these quantitative (“quant”) investment strategies seek to generate active returns by using rules-based factors (like value, momentum and quality) that identify persistent quantifiable drivers of excess returns. These quant-oriented approaches belong to a new investment category (shown in Exhibit 1) that theoretically sits in between passive index strategies offering market beta and active managers who deliver “alpha” (i.e., excess returns not explained by the overall market’s rate of return, or “beta”).2

Factor investing methods may still be relatively new. However, the underlying asset pricing mechanisms for equities have been studied for decades in academic literature by the likes of Eugene Fama and Kenneth French, who explained equity market excess returns, and the risk factor pioneer Barr Rosenburg. The real sea change came when index shops and quant-oriented managers devised transparent methods to measure and capture these factors. Suddenly, quants could outperform the market in ways that only active managers claimed to achieve. To be clear, factor investing doesn’t identify idiosyncratic mispricing (i.e., alpha)—but it began stepping on active manager toes, nonetheless. Why bother with an army of credit analysts who pore through financial statements, if quantitative programs could achieve the same results with a factor lens?

The arrival of bond factor strategies has sparked some robust exchanges between quants and active bond managers. One quarrel that made headlines started when an outspoken quant alleged in a white paper that most active bond managers offer little in the way of true alpha. Any excess returns, they argued, came mostly from “ passive” exposure to corporate credit risks.3 One active bond heavyweight fired back within its own paper, pointing out in careful detail why the accusation was hogwash.4 Skillful security selection from seasoned credit analysts still matters, thank you very much.

QUANTS EXTENDING INTO ACTIVE MANAGER RETURNSExhibit 1: Factors sit between beta and alpha

Exhibit 1: Factors sit between beta and alpha

Source: Franklin Templeton, for illustrative purposes only.

Factor investing, however, isn’t the only quantitative trend bubbling up in fixed income. Among active managers and hedge funds, whose primary mission is generating alpha, an explosion of big data and machine learning algorithms has ushered in a new investing paradigm that some call the “Fourth Industrial Revolution.”If machines can drive cars and translate human speech, then algorithms can surely pinpoint market signals and investable opportunities that traditional active managers might miss. In this machine vs. human scenario, some hedge funds now claim in marketing pitches that data science and machine learning models hold the keys to unlocking real alpha and effectively managing risks.6

It’s worth stating upfront that we think this quant vs. active and machine vs. human debate sets up a false dichotomy. Look under the hood of many bond factor strategies or quant-oriented hedge funds, and you’ll find a deep roster of classically trained economists, portfolio managers and traders—all steeped in the fundamental laws of economics. Machine learning algorithms cannot entirely replace human intuition. We believe that sophisticated models, if not properly guided by professionals with specialized financial expertise, can lead to erroneous conclusions. Fundamental credit analysts still uncover valuable insights that factor-based equations gloss over and don’t understand. A purely quantitative approach (we don’t think they exist in all practicality) is no match for navigating highly dynamic capital markets and the economic force of relentless profit maximization.7

Factors aren’t foolproof

So, why do some factor-based managers shun “active investing approaches” in their marketing pitches? It’s unclear, especially given recent evidence that factors alone aren’t foolproof solutions. Case in point: factor-based equity strategies are suffering from “terrible” performance in 2019—a blunt confession from a quant pioneer at Morningstar’s annual mutual fund conference (this manager plans to “stick like grim death” to his beliefs in factor investing).8

Why the bad performance? Amid the late-cycle market gyrations of 2019, equity factors like value and momentum that historically moved in opposite directions began moving in unison. Quantitative managers long recognized that systematic factors are sensitive to macroeconomic forces; individually, they can underperform for long stretches of time. Building non-correlated multi-factor strategies theoretically alleviated this problem. These diversified factor strategies appeared to work—until they didn’t. Anxious equity investors were told to remain calm and sit tight; this mercurial environment will be short-lived.

Some managers (ourselves included) think static multi-factor exposures are partly to blame. Factor exposures should fluctuate dynamically in response to shifting macroeconomic forces like the rate of unemployment or the overall credit quality of corporate debt markets. Others pound the table and proclaim this poor showing is proof that skilled security selection trumps factors. We think there’s another suggestion: why not capture unique insights from traditionally active and factor approaches at the same time?

Data scientists use the term “ensemble methods” to describe this process of combining different views from multiple algorithms to improve predictive accuracy. Combining active fundamental and quantitative perspectives is also called “orthogonal thinking”—a term used in science to describe a process where unique insights are discovered by drawing on seemingly unrelated perspectives. This orthogonal process was a crucial element, for example, in the discovery of the human genetic code when physicists moved into the field of biology.

Machine learning to the rescue?

If factor-based strategies aren’t foolproof solutions, does machine learning offer a better way to generate alpha? The headlines coming out of tech hubs like Silicon Valley and New York City in the United States and Tel Aviv in Israel tell us data-driven algorithms are accomplishing the unthinkable. Machines now have computer vision, pack groceries in warehouses, drive cars, predict your creditworthiness or even how you’ll vote.

In some areas, machines really do have the upper hand. Who or what else but a machine can identify cats in a stack of 20 million pictures with 95% accuracy in less than five minutes?9 In other areas, humans still prevail. Consider a commuter successfully navigating his or her car through rush hour traffic while holding a work conversation (something computers can’t do) and a crying baby in the back seat. Meanwhile, self-driving cars are still crashing into stationary objects and can become disoriented in the rain.

Now, step out of that car and ask yourself this: how could algorithms navigate incredibly complex and dynamic capital markets that are overflowing with signals and noise? It turns out algorithms are hard-pressed to complete basic tasks inside the noisy environment of finance, where signals are weak and frequently transitory. Scientists describe these environments as having low signal-to-noise ratios.

But there’s good news. Recently published research shows that algorithms can be highly effective when paired with skilled investment professionals (economists and portfolio managers).10 By themselves, machines have trouble anticipating the complicated human responses of politicians and central bankers that can drive market regime changes. However, when operating within the framework of an economist’s hypothesis, algorithms can forecast expected returns with much-welcomed precision—far better than traditional statistical methods, where forecasts remain deeply shrouded in approximation and estimation errors.

Actively guiding quantitative insights

We think the future of fixed income investing requires moving beyond the active vs. quant stalemate. The ideal investment process starts with two familiar dimensions—top-down macroeconomic research and bottom-up fundamental security analysis.

The reasons for this division of labor are relatively straightforward: the performance of nearly every fixed income security (outside sovereign bonds) is influenced by unique mixtures of macroeconomic fundamentals—like inflation and stages of the credit cycle—and sources of bottom-up mispricing tied to individual credit issuers. The skills of a trained economist are different from a credit analyst who specializes in the micro economy of an industry. It takes both viewpoints to capture excess returns, as shown in Exhibit 2.

Two additional dimensions—active fundamental and quantitative science—are also necessary for alpha generation, also illustrated in Exhibit 2. Factor based security selection and machine learning techniques bring new skills and fresh insights to fixed income investing. The goal of combining active with quantitative views, however, isn’t to mix them inside a portfolio like a kitchen blender.

COMBINING ACTIVE AND QUANTITATIVE VIEWSExhibit 2: Model portfolios incorporate ideas across four dimensions

Exhibit 2: Model portfolios incorporate ideas across four dimensions

Source: Franklin Templeton, for illustrative purposes only.

In some instances, quantitative techniques can sharpen fundamental insights with greater precision. Other times, they can challenge assumptions made by fundamental credit analysts and portfolio managers. Through discussion, quantitative views might lead fundamental analysts toward a different conclusion or reconfirm their original hypothesis after a healthy debate and deeper inspection.

It’s been our experience that quantitative methods ultimately work best when combined with “traditional” human insights derived from the academic disciplines of macroeconomics and fundamental security analysis. Yes, machines offer much-needed precision in predicting asset prices, but we still need deep human expertise to make sense of market complexity. Machines aren’t good at assessing turning points in the business cycle or anticipating crowding behaviors from profit-driven traders.

In the next section, we explore top-down macroeconomic research and bottom-up security selection, as shown in Exhibit 3. In terms of top-down, we discuss the process of transforming macroeconomic research into precise return forecasts with algorithms. These serve as a bridge for communicating macroeconomic views across sector teams who speak different languages. In terms of bottom-up, we compare factor-driven insights with fundamental credit analysis. Although factors offer precision across a breadth of securities, deep credit research is still critical for assessing risk premia.

BRINGING FOUR DIMENSIONS TO THE ALPHA-GENERATING PROCESSExhibit 3: Franklin Templeton active quant investment process

Exhibit 3: Franklin Templeton active quant investment process

Source: Franklin Templeton, for illustrative purposes only.

Active quant top down

Anchored in macroeconomics

Developing a well-informed macroeconomic outlook is critical to identifying what academic finance refers to as “risk premia” (i.e., expected returns in excess of the risk-free rate). This type of research is typically the cornerstone of the investment process for active fixed income managers and quant-oriented hedge funds who are paid to deliver alpha. Given the complexity and dynamic nature of capital markets, experienced economists are critical to understanding how a myriad of economic variables can shape expected asset prices.

The sets of signals a macroeconomic team uses to generate an outlook are generally quite diverse and driven by an understanding that business cycles are themselves an aggregation of sub-cycles that drive growth (e.g., personal consumption, residential and non-residential loans, industrial production, services, external demand, etc.). These sub-cycles are interdependent and jointly drive monetary and fiscal policy feedback loops, which in turn impact asset prices.

TRANSFORMING A MACRO OUTLOOK INTO SPREAD FORECASTSExhibit 4: Example macro variables feed our regression tree algorithms

Exhibit 4: Example macro variables feed our regression tree algorithms
Source: Franklin Templeton, for illustrative purposes only.
*This hypothetical example is not a prediction or projection of any investment or investment strategy's performance. It is a hypothetical illustration intended solely to provide insight into how securities are analyzed.

The science of translation

Forecasting sector returns has traditionally been a laborious exercise rife with measurement errors. Analysts typically use statistical techniques like mean reversion (a theory that asset prices eventually revert to long-term averages) or smoothing techniques like moving averages to extrapolate price trends. Some analysts prefer looking backward in time to cherry-pick a previous environment they think best matches the current market regime. All these methods are prone to subjectivity and the likelihood that bond returns may take longer than expected to return to average. Luckily, data scientists have found a better way.

A recently published study on machine learning forecasting and asset prices reflects our own experience modeling expected sector returns (i.e., spreads) using proprietary algorithms. The researchers find that algorithmic models—particularly regression trees that accommodate complex non-linear relationships between multiple variables—“unambiguously improve return prediction” over traditional approaches11 and can also improve portfolio Sharpe ratios.12 A key strength of algorithmic regression trees is their ability to logically capture complex interactions between multiple variables—relationships that human analysts typically find difficult to map out, even with substantial effort and time.

The Franklin Templeton active quant process starts with a macroeconomic outlook provided by trained economists. In the case of our macroeconomic team, it provides our data science team with a range of economic forecasts across variables such as oil prices, exchange rates, equities, interest rates and inflation related instruments. The data science team translates these views into standardized macro variables that feed into the regression tree algorithm, as shown in Exhibit 4.

The output from the regime modeling algorithms effectively translates the macroeconomic team’s outlook into sector views that each sector team can understand. The algorithm serves as a bridge, facilitating discussions among the team to compare and debate sector forecasts. The outcome of these in-depth discussions (i.e., reconciliation process) is summarized in writing each quarter, and available to institutional investors.

We then translate the reconciled sector views into 12-month expected returns across the global multi-sector fixed income universe. This forms the basis of our proprietary sector allocation process. Determining the optimal mix of weights starts by acknowledging spread uncertainties, which are calculated by using the volatility and correlations of a covariance matrix and our market regime models. The results point to an ideal allocation or “starting point” along the efficient allocation frontier. Next, by analyzing all the mandated portfolio constraints, the data science team generates a range of allocations that all fall within the “neighborhood” of the ideal starting point—a region that still meets the portfolio’s risk and return parameters.

As illustrated in Exhibit 5, this optimization process gives portfolio managers the flexibility to allocate across allocation bands. In this hypothetical example, the portfolio managers chose to underweight US investment grade credit by the maximum allowed underweight, while overweighting bank loans and credit risk transfers. This process ensures that risk premia in our portfolios are efficiently allocated across the team’s highest conviction views, and done so in a manner that’s consistent, repeatable and designed for accountability.13

MANAGERS HAVE THE FLEXIBILITY TO ALLOCATE WITHIN BANDSExhibit 5: Portfolio allocation parameters

Exhibit 5: Portfolio allocation parameters

Source: Franklin Templeton, for illustrative purposes only.

Active quant bottom up

Removing blind spots

When you boil down the benefits of bond factor strategies, the standard marketing pitch usually goes something like this: factor-based bond strategies systematically implement investment ideas (using factors like value or momentum) without taking on risks that won’t be compensated (i.e., excessive credit or interest-rate risks). The claim implies that active managers may be doing the opposite—implementing ideas unsystematically (perhaps sloppily), which leads to risk exposures that won’t pay off.

We agree that factor-based strategies offer some advantages, including the ability to methodically (and tirelessly) analyze a wide breadth of securities using precise measures that aren’t subject to bias. But factors also have a significant blind spot: they can’t see what’s causing bond spreads to widen or tighten. On the surface, wide spreads might look attractive to a factor equation, but the equation might be entirely unaware that long-term storm clouds are signaling caution. This blindness can be risky if not supplemented with fundamental research from a seasoned credit analyst—something quants might call “alternative data.”14

It’s the job of the credit analyst to understand how both macroeconomic and microeconomic mechanisms can drive asset prices and explain what’s potentially in store for investors. The credit analyst brings a wealth of information to bear on his or her analysis, from the intricacies of a corporate business model and the peculiar genius (or folly) of a management team, to environmental, social and governance issues.

In this simplified illustration (see Exhibit 6) we’ve captured how Franklin Templeton’s active quant process brings together quant factor-based security rankings and active fundamental credit recommendations into prioritized, potential buy and sell lists at the security level.

COMBINING ACTIVE AND QUANT RECOMMENDATIONSExhibit 6: Using a Punnett Square to identify our target portfolio

Exhibit 6: Using a Punnett Square to identify our target portfolio

Source: Franklin Templeton, for illustrative purposes only.

It’s important to understand the factor models and credit analysts operate independently from each other, ensuring each team’s views remain their own. At the industry level, the “best ideas” from each side are presented in formalized “reconciliation” meetings where credit analysts, portfolio managers and the data team discuss and debate why and how industry views are either synchronized or opposed. Opposing views are entirely welcome—neither quant factors or credit analysts are infallible—and typically lead to deeper analysis and discussion before a resolution is made. We explore this security reconciliation process in the following case study.

Taken together, the active quant process combines potential buy and sell lists with frank discussions and analysis.

The goal is to populate the portfolio with the team’s highest conviction (i.e., highly scrutinized) securities within an industry that also satisfy other risk and strategy-specific constraints.

Factor-based security rankings

Factor-based investing initially became popular as a stock selection strategy by identifying broad, persistent drivers of excess return through quantifiable factors that historically earned positive long-run results. Similar to equity factors, the factor styles for corporate bonds with the longest track records include value, momentum and quality factors. As we outline in the Factor based security calculations section below, each of these factors is grounded in commonly observed market dynamics, such as behavioral biases and structural impediments (rules and restrictions) that create opportunities that fundamental factor investors can exploit.

Historically, factor-based bond managers combine multiple factors into a diversified strategy to help mitigate underperformance of any single factor. All bonds are sensitive to macroeconomic changes and therefore can underperform for stretches of time. Because factor styles like value and quality tend to have low correlations, maintaining fixed exposures to both theoretically reduces the length of underperformance regardless of shifting macro environments.

But, what if you could implicitly forecast the near-term credit cycle? Then a machine learning algorithm could dynamically optimize factor weights according to the expected macro environment, with a goal of reducing potential drawdowns and increasing a portfolio’s Sharpe ratio.

Our algorithmic factor tilts

To forecast the credit environment, our data scientists programmed a gradient- boosting algorithm to incorporate a series of macroeconomic variables such as the unemployment rate, the US Federal Reserve’s balance sheet, and the credit quality of the investment grade and high-yield bond markets. Based on these combined variables, the algorithm predicts the future relative performance of our six style factors spanning the value, momentum and quality categories.

In back-testing, the algorithm dynamically adjusted factor exposures during the global financial crisis—decreasing exposure to Equity Momentum factors and increasing exposure to bonds with lower probabilities of default as measured by Leverage factors and Coverage factors—coverage calculates a company’s ability to pay its interest expenses.

As shown in Exhibit 7, the algorithm began in July 2007 with 76% of factor exposures in Value, with a remaining 21% in Equity Momentum and just 3% in Quality issuers. For historical context, four months later, in October 2007, the Dow Jones Industrial Average stock index peaked at over 14,000 points.

As economic conditions worsened the following year, the algorithm incrementally increased exposure to Quality factors while decreasing exposure to Equity Momentum and Value. Following the dramatic collapse of Lehman Brothers in September 2008, which triggered a global panic, the algorithm increased Quality exposures to 41% by November 2008, with the remaining 52% in Value and just 6% in Equity Momentum.

Instead of a static buy-and-hold approach, our gradient-boosting algorithm shifts factor exposures dynamically to better match overarching macroeconomic environments.

AIMING TO REDUCE DRAWDOWNSExhibit 7: Shifting into quality during the global financial crisis (July 2007–July 2010)

Exhibit 7: Shifting into quality during the global financial crisis (July 2007–July 2010)

Source: Franklin Templeton, for illustrative purposes only.


Four dimensional chess

In today’s rapidly evolving investment landscape, the ability to potentially deliver more consistent excess returns has seen profound changes in the quantitative tools and techniques available to institutional fixed income managers. Outspoken quants who championed the arrival of factor-based strategies are challenging the status quo—daring active managers to prove their worth.

Many active heavyweights are more than ready (thrilled in fact) to meet this challenge, with some getting their arms around big data and machine learning techniques to sharpen their edge. By incorporating data science alongside human insights, a simpler two-dimensional process of top-down and bottom-up analysis has morphed into four-dimensional chess that incorporates fundamental research and quantitative science.

Some managers claim that quants have the upper hand given today’s digital technologies. That isn’t how things are shaping up in practice, however. Algorithms can’t drive themselves in noisy financial environments nor operate successfully without human intuition. Data scientists who lack financial expertise and intuition often don’t produce desired investment results.

In the end, the most important skill sets in fixed income remain the ability of trained professionals to explain the underlying economic mechanisms that drive market regimes and the signals that data science can track and analyze. The future of fixed income has already arrived—it lies in successfully marrying quantitative science with fundamental based active management.

Franklin Templeton Fixed Income Group has engineered a seamless active quant approach—where portfolio managers, analysts, traders and data scientists work as one team to create a synergistic loop between quantitative and fundamental analysis. We believe marrying our data science and fundamental expertise gives us the insights and competitive edge to navigate challenging investment environments and serve our clients better.

ENDNOTES

  1. Source: Baker, S. “Factor-based investing arrives for fixed income,” Pension & Investments, 16 October 2017.

  2. Source: J. Bender, R. Briand, D. Melas, and R. Subramanian. “Foundations of Factor Investing,” MSCI Research Insight, December 2013.

  3. Source: AQR, “The Illusion of Active Fixed Income Alpha,” Alternative Thinking, 17 December 2018.

  4. Source: J. Baz, M. Devarajan, M. Hajo, and R. Mattu. “When Alpha Meets Beta: Managing Unintended Risk in Active Fixed Income,” PIMCO, Quantitative Research, May 2018.

  5. Source: M. Kolanovic and R. Krishnamachari, “Big Data and AI Strategies—Machine Learning and Alternative Data Approach to Investing,” J.P. Morgan, May 2017.

  6. Source: Two Sigma, “Our Approach,” September 2019.

  7. Source: Fama, E. “Efficient Capital Markets: A Review of Theory and Empirical Work,” Journal of Finance, Vol. 25, No. 2, Papers and Proceedings of the Twenty-Eighth Annual Meeting of the American Finance Association, New York, N.Y., December 28–30, 1969 (May 1970).

  8. Source: Lee, J. “AQR’s Asness Is Right. It’s a ‘Crappy’ Time for Factor Investing,” Bloomberg Markets, 15 May 2019.

  9. Brownlee, J. “How to Classify Photos of Dogs and Cats (with 97% accuracy),” Deep Learning for Computer Vision, 3 October 2019.

  10. Source: S. Gu, B. Kelly and D. Xiu, “Empirical Asset Pricing via Machine Learning,” Chicago Booth Research Paper No. 1818-04. 13 September 2019.

  11. There can be no assurance that any model, whether algorithmic, traditional, or otherwise, can predict return.

  12. Source: S. Gu, B. Kelly and D. Xiu, “Empirical Asset Pricing via Machine Learning,” Chicago Booth Research Paper No. 1818-04. 13 September 2019. The Sharpe ratio is the average return earned in excess of the risk-free rate per unit of volatility or total risk.

  13. There can be no assurance that adopting this optimization process will have any impact on investment outcomes, or that it will result in profits or that it will minimize losses.

  14. Including alternative data in the analysis does not necessarily reduce or eliminate risk.

Franklin Templeton Thinks: Fixed Income Markets highlights the team’s ongoing analysis of global economic trends, market cycles and bottom-up sector insights. Each quarterly issue spotlights the team’s thinking on different macro forces, and particular sector views that drive our investment process.

Primary contributors to this issue

Patrick Klein, Ph.D. Portfolio Manager,
Multi-Sector Strategies

Thomas Runkel, CFA Portfolio Manager

Pururav Thoutireddy, Ph.D. Quantitative Research Analyst



Editorial Review

Sonal Desai, Ph.D. Chief Investment Officer,
Portfolio Manager

David Yuen, CFA, FRM Director of Quantitative Strategy,
Portfolio Manager

John Beck Director of Fixed Income, London

David Zahn, CFA, FRM Head of European Fixed Income,
Portfolio Manager

The Franklin Templeton Fixed Income Group

Franklin Templeton has been among the first to actively invest in many sectors of the fixed income markets as they have evolved—covering corporate credit, mortgage-based securities, asset-backed securities and municipal bonds since the 1970s, international fixed income since the 1980s and bank loans since the early 2000s. Over 170 investment professionals globally support the portfolio managers, who oversee more than US$169 billion in assets under management. Being part of an established investment group at Franklin Templeton gives the portfolio managers access to experts across different areas of the fixed income market, helping them to diversify opportunities and risks across multiple sectors.

Our global reach through Franklin Templeton Investments provides access to additional research, trading, and risk management resources. Portfolio managers have opportunities to exchange insights with other investment groups, and collaborate with an independent risk team that regularly examines risk analytics to help identify and address areas of excessive risk exposure within our portfolios.