Corporate Bond Selection: Momentum, Carry, and Machine Learning (Part 2)

Previous: Picking Corporate Bonds: Credit-Sensitive Security Selection (Part 1)

In Part 1, we covered value signals and default models for corporate bonds. Now let’s get into the other three investment themes: momentum, carry, and defensive. Then we’ll see what happens when you combine them all. And yes, we’ll talk about machine learning too.

Momentum in Corporate Bonds

Momentum in stocks is a well-known idea. Buy recent winners, sell recent losers. It works in corporate bonds too, but with a twist.

Richardson uses two types of momentum signals. The first is “own” momentum. You look at a bond’s credit excess return over the past six months. Bonds that did well recently tend to keep doing well. Bonds that did poorly tend to keep doing poorly.

The second is “related” momentum. This one is interesting. Instead of looking at the bond itself, you look at the stock price of the company that issued the bond. If a company’s stock has been rallying for six months, its bonds tend to do well going forward too. This idea goes back to research by Kwan in 1996. The catch is that this only works for companies with publicly traded stock. Private issuers don’t have stock prices to track.

One detail worth noting. In stock momentum strategies, people usually skip the most recent month to avoid bid-ask bounce effects. You don’t need to do that when using equity momentum to pick bonds. The cross-market contamination isn’t really an issue. If anything, the most recent equity returns are the most useful for predicting credit performance.

And momentum doesn’t stop there. You can build momentum signals from changes in fundamental data, changes in default probability forecasts, or even from supply-chain-linked firms. If Company A is a major supplier to Company B and Company A’s stock is ripping higher, that might tell you something about Company B’s bonds. Momentum can become a very broad theme for systematic credit investors.

Carry: Getting Paid to Hold

Carry in corporate bonds is the same idea as carry in government bonds. You’re trying to figure out how much you’d earn just by holding the bond, assuming nothing changes about the issuer’s credit quality.

A full carry calculation requires building an issuer-specific credit term structure. That’s a lot of work. So Richardson keeps it simple here and uses the option-adjusted spread (OAS) as the carry measure. Higher spread means higher carry. You’re getting paid more to hold that bond.

Simple as that.

Defensive: Low Risk and Quality

The defensive theme has two parts.

The first part is about low risk. The idea is that lower-risk bonds tend to deliver better risk-adjusted returns than you’d expect. Why? Partly because of leverage aversion. Many investors can’t or won’t use leverage, so they reach for riskier bonds to boost returns. That leaves the safer bonds underpriced on a risk-adjusted basis.

Richardson uses spread duration as the low-risk measure. In credit markets, a bond’s beta can be approximated by DTS (duration times spread). But DTS mixes two things together. Spread duration is the risk component, and spread is the carry component. Those two have opposite effects on expected returns. So Richardson separates them. Spread duration goes into the defensive bucket. Spread goes into the carry bucket. Clean and logical.

The second part of defensive is quality. Richardson uses two measures here: market leverage (prefer companies with less debt) and gross profitability (prefer companies making more money). Lower leverage and higher profitability both point toward healthier companies whose bonds should hold up better.

Putting It All Together: Performance

Now the fun part. How do these signals actually perform?

Richardson tests all four themes (value, momentum, carry, and defensive) on a combined universe of US investment-grade and high-yield bonds from 1997 to 2020. Each month, bonds are ranked on each signal. Academic long-short portfolios go long the top 20% and short the bottom 20%.

The results are strong. Sharpe ratios range from 0.98 for momentum to 1.71 for value. Carry is the weakest on its own. But the real magic happens when you combine all four themes with equal weighting. Because the signals have low correlations with each other, the combination portfolio is better than any single theme.

How low are those correlations? Low enough that the combined signal has a regression intercept t-statistic of 12.7 after controlling for traditional market risk premia and well-known equity factors. That’s absurdly high. Richardson himself calls it “too good to be true,” but quickly adds that these are academic portfolios. They ignore transaction costs, liquidity constraints, and all the messy real-world stuff.

One key finding: credit value investing is uncorrelated with equity value investing. That might seem weird since both are about buying cheap assets. But the measures are different (spread-to-default-probability vs. book-to-price), the companies are different (some have public debt but no equity, and vice versa), and the time horizons are different. This matters because equity value investing had a terrible decade. Credit value didn’t share that pain.

When you separate IG and HY universes and use better portfolio construction (sector demeaning, beta neutralizing), the results get even more interesting. Momentum works better in high yield. Carry works better in investment grade. Safer bonds have more spread mean reversion, which favors carry. Riskier bonds have more continuation in performance, which favors momentum.

The Liquidity Premium Myth

Here’s a popular belief: less liquid bonds should offer higher returns to compensate investors for the difficulty of trading them. Sounds logical, right?

Richardson points to research by Palhares and Richardson (2019) that tested this idea six different ways. They measured liquidity using issue size, bid-ask spreads, market impact, daily trading volume, percentage of no-trading days, and bond age.

The cleanest test looks at different bonds from the same issuer. If one bond is liquid and another is illiquid, any spread difference can’t be explained by credit risk since it’s the same company. The result? Less liquid bonds do have slightly wider spreads. But when you look at actual returns, there’s no statistically significant evidence that holding less liquid bonds pays off.

This is a big deal for real-world investors. If you can’t find a liquidity premium even in a frictionless academic setting, the result after paying transaction costs will probably be negative. And holding illiquid bonds in a fund with daily redemptions creates a whole other problem.

Machine Learning for Default Prediction

Richardson is refreshingly honest about machine learning. He says he’s “no expert” but has “used them successfully.” He’s also skeptical of the hype.

His practical take: corporate bonds don’t generate enough trading data for ML to help with execution or trading. But ML can help with default forecasting, where you have 50+ years of data covering thousands of companies.

The specific technique he highlights is random forests. A pruned decision tree starts with all firms and splits them based on whichever variable best separates the bankruptcies from the survivors. The first split? Leverage (asset value divided by debt). High leverage firms have a 4.8% bankruptcy probability vs. 0.2% for low leverage firms. The second split uses volatility.

This is elegant because the tree naturally discovers the interaction between leverage and volatility that sits at the heart of structural credit models. But it does so from the data, without assuming the model upfront. It can also find nonlinear relationships and interactions that traditional models miss.

The risk of overfitting is real, of course. But cross-validation and randomization across many trees (that’s why it’s called a random forest) can help keep things honest.

The Communication Challenge

One last thing Richardson emphasizes: systematic investors need to get better at explaining what they do. Asset owners vary in how much they want to understand, but trust matters. If you can’t clearly explain your investment process before and after, clients will leave. Even if your returns are fine.

A systematic manager will never have a story for every single bond in the portfolio. But you can build visual tools that let clients look through the portfolio and see why positions exist and what’s driving returns. It’s a journey, Richardson says, that you and your investors take together.

Key Takeaways

Momentum, carry, and defensive signals all work for corporate bond selection. But they work differently depending on whether you’re in investment grade or high yield. Momentum shines in riskier credits. Carry shines in safer ones.

The liquidity premium in corporate bonds is mostly a myth. Don’t hold illiquid bonds hoping to earn extra returns.

Machine learning can genuinely help with default prediction, but it’s not a magic solution for everything in credit investing. Start with good data and clear economic intuition.

And above all, combining multiple uncorrelated signals is where the real power lies. Value, momentum, carry, and defensive together produce something much better than any single theme alone.

Next: Emerging Market Bonds: Hard Currency Security Selection

Book: Systematic Fixed Income: An Investor’s Guide by Scott Richardson, Ph.D. Published by John Wiley & Sons, 2022. ISBN: 9781119900139.