The -0.14 bps ex-Covid mean is too thin a basis for a trade without evidence of how fast the gap closes. Heidorn, Klaus and Mazzalupi have assembled a useful map of the European CDS-bond basis by sector. Their headline claim, that the basis "is fundamentally mean-reverting over longer time horizons," goes further than the charts and summary statistics can carry. The paper reports no P&L or net-of-cost test.
The matched pairs
The basis takes the 5-year single-name CDS spread and subtracts the Z-spread of a matched cash bond, measured over the €STR OIS curve. A negative reading means the bond pays more spread than protection costs. The negative-basis position holds the bond and buys protection to collect that gap.
The universe covers every constituent of iTraxx Europe 5Y from Series 33 to Series 42: ten six-month series with daily observations from March 20, 2020 to March 20, 2025. The authors match each CDS by hand to one senior unsecured bond from the same legal entity. It must have at least €450 million outstanding, and its maturity must be as close to the CDS as an available bond allows. There is no interpolation. Most bonds are EUR; a few are GBP or CHF. This produces 1,172 matched bonds from 1,250 CDS names and 32,877 daily observations. The maturity gap averages -0.18 years, with a range from -0.61 to +0.24. The paper reports index and Markit's five-sector moments, then draws on conversations with traders to interpret them.
Across the full sample, the mean basis is -4.17 bps, the median 1.50 bps and the standard deviation 20.03 bps. The -102.67 bps minimum falls in the Covid series. CDS averages 67.87 bps; cash averages 72.04 bps.
What changes when S33 goes?
The case for mean reversion draws heavily on a second table, which removes S33 to strip out Covid. The mean becomes -0.14 bps, the standard deviation falls to 15.20 bps and the minimum rises to -31.55 bps. The authors describe the exclusion as a way to "observe basis behavior under more regular circumstances." In the conclusion, the -0.14 bps mean is used as "reaffirming the long-run no arbitrage relationship."
S33 contains the -102.67 bps minimum. Removing it after seeing the data gives a picture of a calmer sample, yet the conclusion treats that picture as confirmation. The median also moves farther from zero, from 1.50 bps to 4.12 bps, even as the mean approaches it. The authors say bonds generally trade slightly tighter than CDS. On a typical ex-Covid day, then, the basis is about +4 bps. Fewer, larger negative days pull its mean close to zero. They make the same point about the full sample, and it remains true ex-Covid. A series that spends most days positive and turns negative in stress has a different trading character from one that repeatedly crosses zero.
A sample mean says little about the path back.
We found no unit-root test, half-life estimate or cointegration test in the paper. Engle and Granger appear only in its list of consulted literature. A trader needs a speed of reversion to choose a holding period; without a half-life, there is no holding period to size against.
Financials against the rest
The sector tables are the paper's strongest evidence. Financials average -18.38 bps (median -17.84, minimum -141.31). Their cash spread is 87.22 bps against 68.84 bps for CDS. Autos & Industrials average +6.67 bps (median 10.65), and Energy averages +3.80 bps. Consumers average -8.02 bps; TMT (technology, media and telecoms) averages -4.70 bps. Autos & Industrials account for roughly 28 to 29% of names. Financials account for 24%, with 30 names in every series. These are the two largest blocks, so the -4.17 bps index mean nets sectors with opposite signs. The authors' charts show the effect during S37 to S39: Energy and TMT move against Financials while the index basis stays roughly flat.
Financials, the paper says, are "heavily influenced by Credit Suisse turmoil." Figure 8 shows much the same shape after removing Credit Suisse, at a smaller magnitude. As the authors put it, "the movement of the Financials basis is essentially the same but with a smaller magnitude." They quantify the CS effect only at index level, at 1.5 bps on average and 13 bps at most over S37 to S39. There is no Financials-ex-CS mean.
The chosen Credit Suisse window is wide. The authors begin the stress with the Suisse Secrets data leak in early 2022; the collapse came in March 2023. S37 to S39 run from March 2022 to September 2023. Averaging 1.5 bps across eighteen months dilutes the weeks that mattered.
TMT raises a separate sample-size concern. It has 16 to 20 names and the highest basis volatility, 27.00 bps, compared with 19.11 for CDS and 24.34 bps for cash. We could not find how the sector or index series are aggregated, whether equal-weighted or otherwise. That choice determines how much one issuer can move a 17-name sector.
Who could book the gap?
The authors acknowledge that "this calculation doesn't take into consideration the bid-ask spread in both markets, which could create a minor bias." Their conclusion says execution costs and bid-ask spreads "could partly bias the basis levels." I disagree with "minor." Practitioners interviewed for the paper say executing 10 to 20 million in single-name CDS is standard, while selling 20 to 30 million of a bond at a reasonable spread is much harder. Those observations concern stress, precisely when the basis is most negative. A round trip crosses four half-spreads, two on each leg. The paper reports none of them.
The interviews call negative-basis bottoms "optimal entries for specific funds focusing on arbitrage opportunities." A cash illiquidity premium, the paper says, "could allow the investor to lock in a great basis, ensuring risk free return." No P&L accompanies that claim. The conclusion instead lists "arbitrage strategies employed by fund managers and market makers" as future work. Bond funding and repo enter the paper's analysis only through the practitioner section. Its spread measure is the Z-spread, though the authors acknowledge that the market focuses on the asset swap spread. A positive net-of-cost backtest on the 1,172-bond panel, entering below a fixed threshold and charging funding on the bond leg, would change my view.
We could not run that test ourselves. We have neither single-name CDS spreads nor cash corporate bond spreads, and the trade depends on their difference. ETF or futures proxies would fail to preserve the basis.
The sector tables show where European bank bonds stood against their CDS for five years. The mean-reversion claim remains a hypothesis the authors have the data to test.