Key takeaways
- Across the 1866-2008 sample, credit spreads averaged about 153 basis points a year against roughly 75 basis points of realised default losses, leaving a credit risk premium near 80 basis points.
- On 1997-2003 averages, expected default loss explained 11.8% of a 3-5 year BBB spread and 57.9% of a B spread. The premium is mostly a high-grade phenomenon.
- Senior unsecured recovery averaged 38.4% over 1982-2006 but reached 58.3% in 2006. On a 4.90% default rate, that swing moves annual credit loss from 3.02% to 2.04%.
- In October 2008 the Federal Reserve's excess bond premium was 3.40 percentage points of a 7.14-point credit spread, so 47.6% of the spread was risk appetite alone.
- Sterling high-yield illiquidity premia rose 121 basis points between early August and end-November 2007, against 69 for dollar and 57 for euro spreads.
Roughly half the spread paid for defaults that happened, and half was premium
A corporate bond yields more than a government bond of the same maturity. That gap is the credit spread, and the question worth asking is what it's compensating you for, and how much of it a holder actually keeps.
The longest answer available comes from Kay Giesecke, Francis Longstaff, Stephen Schaefer and Ilya Strebulaev, who assembled a corporate bond default series running from 1866 to 2008. Over that stretch the average annual default rate was about 1.50%. Apply a 50% loss rate, meaning a defaulted bond returns half its face value, and average annual credit losses come to roughly 75 basis points. Credit spreads over the same period averaged about 153 basis points.
So about 49% of the spread was matched by estimated losses on the defaults that actually happened. The rest, about 80 basis points a year, is what the authors call the credit risk premium. Over the 1866-2008 sample that's what a credit allocation was really being paid, before any fund fee, for agreeing to hold an asset that can default.
Two things follow. The extra yield is not free money, because part of it is a bill that arrives later. And it isn't a pure loss estimate either, because most of the time the bill was smaller than the payment.
A spread has four jobs, and expected default loss is the smallest of them
The Bank of England's 2007 study by Lewis Webber and Rohan Churm sets out the anatomy cleanly. Part of the spread compensates for expected default loss. Another part compensates risk-averse investors "for their exposure to unexpected default losses". A third part is a relative illiquidity premium, because the government bond market is normally easier to trade than the corporate one. What's left is a residual: tax treatment, call options, collateral eligibility.
Put numbers on it and the ranking is not what most people expect. The Bank for International Settlements paper by Jeffery Amato and Eli Remolona averaged option-adjusted spreads from January 1997 to August 2003 and set them beside an expected-loss estimate built from a ratings transition matrix and assumed recovery rates. At three to five years, AAA spreads averaged 63.86 basis points against expected losses of 0.18. BBB averaged 170.89 against 20.12. B averaged 691.81 against 400.52.
That BBB row is the one to sit with. The average spread was about 170 basis points and the average yearly loss from default was 20. The gap between spread and expected loss widens from 64 basis points at AAA to 291 basis points at B.
The premium share collapses as credit quality falls
Turn those pairs into percentages and you get the chart above: the share of each spread that expected default loss does not explain. At AAA it's 99.72%. At AA, 97.98%. At A, 97.30%. At BBB, 88.23%. At BB, 65.23%. At B, 42.11%.
Read left to right, that line is the whole argument. Buying a high-grade corporate bond over a government bond is almost entirely a bet on something other than defaults. Buying a B-rated bond is mostly a bet on defaults, with a thinner premium attached. The 1997-2003 sample is one credit cycle, so treat the levels as illustrative and the slope as the finding.
It also explains why credit behaves differently from rate risk. A spread can widen while government yields fall, and the two effects work on your capital independently. If you want the rate side of the arithmetic, our guide to what a 1% rate move costs a bond holder covers the duration mechanics that sit underneath any credit position.
Recovery is the assumption doing the most work, and it moves against you
Expected loss is a product of two numbers: how often issuers default, and how much you get back when they do. The second is the one people wave through.
The BIS calculation assumed recoveries of 68.34% at AAA, 49.42% at BBB and 37.54% at B. Those aren't observations of the future; they're historical averages from Altman and Kishore. Moody's twentieth annual default study puts the volatility on display. Senior unsecured recovery averaged 38.4% over 1982-2006, but it reached 58.3% in 2006, up from 55.5% in 2005.
Run that swing through the default rate and it's large. Moody's speculative-grade default rate averaged 4.90% a year from 1983 to 2006. At a 38.4% recovery, that's an annual credit loss of 3.02%. At a 58.3% recovery, it's 2.04%. The recovery assumption alone is worth 0.98 percentage points of loss on the same default experience — comfortably more than the entire spread on an investment-grade index today.
Worse, recovery isn't independent of defaults. Moody's regressed annual senior unsecured recovery rates on annual corporate default rates and got an R-squared of 0.55. When defaults cluster, recoveries fall, because distressed assets get sold into a market with the same problem. Your loss estimate and your loss experience go wrong in the same direction at the same time. That's not a rounding error in the model, it's the model's weakest joint.
Taxes and trading costs explain more of an investment-grade spread than defaults do
If expected loss is a tenth of a BBB spread, what fills the rest? The BIS paper reproduces two decompositions. Elton and co-authors split a 10-year BBB spread into 34.7% expected loss, 28.4% taxes, 30.0% risk premium and 7.0% other. At AA and five years, the tax share reaches 72.6%.
The tax component is a US artefact. Corporate bonds are taxed at state level there and Treasuries are not, so a corporate yield has to carry the tax. Elton's benchmark rate of 4.875% lets taxes explain 28% to 73% of spreads depending on rating and maturity. Driessen, using different data, puts taxes at 34% to 57%, risk premia between 18% and 52%, and liquidity premia at about 20%.
For a UK reader that matters twice over. A sterling corporate bond fund isn't paying US state tax, so that slice of the American research doesn't transfer. And whatever the American tax wedge was doing, it wasn't compensating anyone for credit. It's the same trap as reading a nominal gilt yield as a real return, which is the distinction our piece on index-linked gilts and the inflation uplift lag works through in detail.
Trading costs are the other quiet slice. Schultz estimated round-trip trading costs in the US corporate bond market at about 27 basis points. On a 97 basis point BBB spread, one round trip eats more than a quarter of a year's compensation.
The premium is a price, not a constant: 3.40 points of it in October 2008
The Federal Reserve publishes a monthly series that does this split in real time. It runs from January 1973 and separates a corporate credit spread into the part explained by measured default risk and a residual the Fed calls the excess bond premium — investor risk appetite, in plain terms.
In October 2008 the credit spread stood at 7.14 percentage points and the excess bond premium at 3.40. That means 47.6% of the spread on offer was compensation for nothing you could measure in the issuers' balance sheets. In March 2020 the spread was 3.82 points with a premium of 1.19. In July 2026 the spread was 0.84 points and the premium was negative, at -0.32.
Across all 643 monthly observations from 1973 to July 2026 the premium averages 0.06 percentage points, which is close to zero by construction. The point isn't the average. It's the range: the same nominal spread can be generous or mean depending on when you buy it, and the difference has nothing to do with the default outlook.
The liquidity slice arrives exactly when selling is hardest
The illiquidity premium is the component that behaves worst under stress. The Bank of England's model attributes it as the residual, and tracked it through 2007. Between early August and the end of November 2007, the illiquidity component of sterling high-yield spreads rose 121 basis points. Dollar spreads rose 69 and euro spreads 57.
Nothing about issuer fundamentals moved that fast. What moved was the willingness to hold something you might not be able to sell. The same model showed the non-credit residual on dollar investment-grade spreads widening by around 30 basis points between the end of 2001 and mid-2002, around the US accounting scandals — a period when the reliability of balance sheets, not their contents, was the question.
If you hold a corporate bond fund, that's the part of the spread you're least likely to collect. It compensates for the option to sell at a fair price, and it's paid out to whoever is still willing to buy when that option is worth the most.
The strongest objection: through credit default swaps, default risk is most of the spread
There's a serious body of work that says the decomposition above understates default. Francis Longstaff, Sanjay Mithal and Eric Neis used credit default swap premia for 68 firms between March 2001 and October 2002 to measure the default component directly. Against the Treasury curve, they found it accounted for 51% of the spread for AAA and AA bonds, 56% at A, 71% at BBB and 83% at BB.
Those numbers look flatly incompatible with an 11.8% expected-loss share at BBB. They aren't, and the reason is the most useful thing in this article. A credit default swap premium is a market price, so it already contains the risk premium — it's what someone charges to take the default risk, not a forecast of how often default happens. The 20 basis point figure is a physical estimate built from historical default frequencies. Both can be right at once: the "default component" of a spread is large, and the part of it that expected losses justify is small.
What the swap-market evidence does settle is the residual. Longstaff and co-authors found a significant non-default component for every firm in their sample, ranging from about 20 to 100 basis points, and it tracked bid-ask spreads and issue size rather than tax rates. The liquidity slice is real, and it is not an artefact of a badly specified model.
Running the same arithmetic on today's spreads
The ICE BofA US Corporate BBB option-adjusted spread closed at 0.97% on 6 August 2026, with the broad investment-grade index at 0.78% and the BB tier at 1.61%. Set those against Moody's average one-year credit loss rates for 1982-2006: 0.107% for Baa issuers, 0.744% for Ba, 3.317% for B.
At BBB, the historical loss rate is 11.0% of the current spread. At BB, 46.2%. The same downward slope from the 1997-2003 data shows up in 2026, on entirely different numbers. Whether a 0.97% spread is generous depends less on your default forecast than on what you think the other 89% is worth to you.
Averages also hide the tail. Moody's speculative-grade default rate averaged 4.90% but ran as high as 10.59% in a single year, and dispersion inside high yield is where that shows up first — a point our analysis of high-yield spreads and credit dispersion takes apart at the rating-bucket level.
What the data cannot tell you
Start with the sources. The current Moody's and S&P annual default studies sit behind subscriber logins, so the recovery and loss figures here come from the Moody's annual study covering 1920 to 2006, which is readable without one. It stops before the 2008 default wave and before 2020, so neither episode is inside any of these averages.
The spread history has thinned too. FRED now carries only 3 years of observations for the ICE BofA option-adjusted spread series, so the modern comparison above is a snapshot rather than a cycle.
Every decomposition here is a model, and the residual absorbs whatever the model failed to specify. Calling that residual "liquidity" is an interpretation, not a measurement. The BIS averages cover 1997 to 2003, one cycle. The swap-market study covers 68 firms over 19 months, most of them investment grade, and it circulated as a working paper. The 1866-2008 series is US, and its 50% loss rate is an assumption applied uniformly across 143 years, not something anyone observed.
Nor do default statistics arrive on the schedule intuition expects. Giesecke and co-authors found the correlation between credit events and NBER business downturns was only about 26%, and that the worst three-year Depression default total of 12.88% for non-financial issuers only barely made their top five credit events. The 1873-1875 railroad crisis took out 36% of the par value of the entire corporate bond market. A sample that hasn't yet seen its worst decade flatters the premium.
Finally, an index spread isn't your spread. You hold a fund, with a fee and a cap-weighted index behind it, and the compensation reaching you is what's left after both.
What would change the conclusion
If recovery rates were stable, the expected-loss slice could be estimated in advance with some confidence, and the honest premium would be smaller than 80 basis points. The R-squared of 0.55 between defaults and recoveries says they aren't stable, and that they fail together.
If spreads adjusted to realised defaults, the premium would be a forecast rather than a price, and a run of clean years would compress it toward expected loss. Giesecke and co-authors tested exactly this over 1866-2008 and reported that credit spreads do not adjust in response to realized default rates. Should that finding break on post-2008 data, the framework here weakens considerably.
If corporate bonds ever traded as cheaply as government bonds, the illiquidity slice would go, and a 121 basis point stress move in sterling high yield would stop being possible. Electronic trading has narrowed the gap; the 2007 episode says it hasn't closed it.
The number to watch isn't the spread. It's the distance between the spread and a loss estimate you'd be willing to defend out loud — at BBB that distance was 88.23% of the spread in 1997-2003, and something close to it today. When that distance is wide, you're mostly being paid to bear uncertainty. When it's narrow, you're mostly being paid to absorb losses, and those are different jobs.