Project Record

Bond–CDS Basis and Stress Strategies

Summarize the Ionitsa project record titled "Bond–CDS Basis and Stress Strategies" for a technical reader.
Cover the problem or research question, implementation or method, evidence or results, and limitations.
Separate facts stated on the page from your own assessment, note anything unclear or unverified, and avoid promotional language.

Primary source: https://ionitsa.com/projects/cds-bond-market-neutral.md
Canonical page: https://ionitsa.com/projects/cds-bond-market-neutral/

Reproducible research into cash-versus-synthetic credit dislocations, beginning with one issuer and expanding into stress events and CDX rolls.

Executive Summary

Cash bonds and credit default swaps express credit risk through different markets. In ordinary periods they often move together. Under funding pressure, ratings changes, bankruptcy rumours, index rolls, or an actual credit event, their prices can separate sharply. That separation may be a trade—or merely the price of financing, liquidity, delivery options, and being unable to transact at the observed mark.

I am building a reproducible research system to tell those cases apart. The project begins with one issuer, one bond, one five-year CDS contract, and an explicit rates hedge. It keeps data preparation, contract mapping, backtesting, and event analysis in a standalone Python project with a Jupyter notebook. The Astro pages are publications derived from that work rather than substitutes for it.

The first case study is Ford around its 2020 fallen-angel episode. The initial implementation includes a strict data contract, CS01-based hedge sizing, next-observation execution, transaction and financing costs, stale-data gates, and a deterministic synthetic stress dataset that tests the pipeline. Real performance conclusions remain gated on licensed historical single-name CDS and institutional bond data.

This demonstrates credit relative-value research, derivatives and cash-market risk decomposition, event-driven backtesting, and the discipline to separate a working research system from an unsupported performance claim.

Research Programme

The project is the durable unit of work. Individual research articles can later be produced from specific experiments without duplicating or obscuring the code that generated them.

The first article is now published: The Discrepancy Is Not the Strategy: A Bond–CDS Basis Stress Test. It reviews the literature, reproduces its principal mechanisms with transparent stylised data, and derives a funding-gated negative-basis strategy.

The follow-up article is When One Name Breaks the Basket: CDX, Defaults and the Index Skew. It explains series, versions, index factors and credit-event settlement, then tests CDX against a PV01-weighted basket of 125 stylised single-name CDS contracts.

The valuation follow-up is Pricing Before Prediction: How to Build and Test a Credit Engine. It adds the tested hazard-curve, CDS cash-flow and bond Z-spread layer needed to replace first-order P&L during large credit moves.

flowchart LR
  A[Vendor and public raw data] --> B[Point-in-time normalization]
  B --> C[Issuer and contract mapping]
  C --> D[Risk and cash-flow engine]
  D --> E[Backtests and event studies]
  E --> F[Auditable result tables]
  F --> G[Astro research articles]

The repository lives under research/cds-bond-basis/. It contains the input contract, Ford case-study configuration, tested Python engine, deterministic demo generator, and starter Jupyter notebook. Licensed raw data, credentials, and generated outputs are deliberately excluded from version control.

Basis Definition

For a bond par-equivalent spread $s^{bond}_t$ and matched CDS par spread $s^{CDS}_t$, define:

A negative basis suggests buying the bond and buying CDS protection. Ignoring frictions, the bond’s excess spread is greater than the protection premium. A positive basis suggests shorting the bond and selling protection. Neither trade is an arbitrage by definition: funding, repo or borrow, accrued amounts, counterparty terms, the cheapest-to-deliver option, and liquidity can explain a persistent basis.

The bond spread must be comparable with CDS. A raw yield spread is generally insufficient because it mixes the credit view with coupon, curve, optionality, and rates effects. The baseline accepts a Z-spread for plumbing tests; production research should prefer a par-equivalent or model-consistent bond spread.

What “Neutral” Means

The previous version of this project matched a bond’s rate DV01 directly to a CDS sensitivity. Those are different risks. The corrected framework separates them.

Let $CS01_b$ be the bond’s change in value for a one-basis-point credit-spread move, and let $RPV01_c$ be the CDS risky present value of one basis point per unit of notional. For bond notional $N_b$, the credit-matched CDS notional is:

For a long bond, $N_c$ is protection bought; for a short bond it is protection sold. The bond still has rates sensitivity. With bond rate DV01 $DV01_b$ and a Treasury hedge instrument worth $DV01_h$ per contract, the first-order hedge is:

This creates first-order credit-spread and rates neutrality under the assumed mapping. It does not remove curve, convexity, volatility, jump-to-default, recovery, liquidity, or model-basis risk.

Backtest Design

The first strategy is intentionally plain. It standardizes the basis against a rolling history, enters outside a pre-declared threshold, and exits after convergence:

The rolling mean and standard deviation end at $t-1$. A signal observed on date $t$ becomes a held position on the next observation. This lag prevents a common same-close look-ahead error.

Daily approximate P&L decomposes into:

The starter strategy engine uses spread-duration approximations so every term remains auditable. The companion Credit Pricing and Model Validation Engine now measures that approximation against dirty-price bond and CDS cash-flow revaluation. An institutional backtest must connect those analytics to the actual contract, recovery and default settlement, and the traded rates hedge.

Stress and Event Studies

The strategy layer and event layer are separate. A profitable-looking full sample can hide a single catastrophic default or depend entirely on one crisis. Each declared event therefore receives a pre-window, event window, and post-window with gross and net attribution.

The Ford configuration begins with the 25 March 2020 S&P downgrade from BBB- to BB+. The event is useful for examining a fallen angel during a market-wide liquidity shock, but it is not clean identification by itself. Results must distinguish the broad COVID credit shock, the ratings action, forced cash-index selling, central-bank intervention, and the CDS market response.

Later experiments can examine:

1. ratings-watch, downgrade, and fallen-angel transitions;
2. bankruptcy rumours versus formal ISDA credit-event decisions;
3. CDX roll inclusion and exclusion using point-in-time constituent files;
4. a credit event’s transition from the full to reduced index;
5. cross-sectional basis dispersion after funding or liquidity shocks.

CDX is the North American CDS index family. Ordinary constituent changes happen at index rolls; after a confirmed credit event, index mechanics can instead move to a reduced version. Those are different event types and must not be collapsed into a generic “exclusion” flag.

Data Plan

Public sources can validate parts of the pipeline, but not the entire historical trade.

The production dataset records timestamps and vendor/as-of metadata. Combining a TRACE execution at one time with a CDS end-of-day composite from another time can manufacture a basis that was never executable.

First Deliverable

The minimum credible real-data study is deliberately narrow:

• one Ford Motor Company bond selected for liquidity and a stable 4–7 year maturity around the start of the sample;
• the legally matched USD senior five-year CDS contract, including reference entity, tier, restructuring clause, and RED identifier;
• daily observations from 2019 through 2021;
• a matched Treasury or swap rates hedge;
• declared March 2020 event windows;
• gross and net P&L, turnover, stale marks, sensitivity analysis, and tail loss.

The bond must not be chosen retrospectively because it produced the best basis. Ford Credit and Ford Motor Company obligations also require an explicit entity and deliverability check; name similarity is not enough.

Failure Conditions

The research should reject—not massage—a result when contract identity cannot be verified, observations are stale through the stress window, bid/ask and borrow erase the apparent edge, the trade depends on same-day unavailable information, or the conclusion disappears under reasonable spread and funding assumptions.

The synthetic notebook proves only that the plumbing, risk signs, lags, and costs behave as intended. It is intentionally incapable of proving historical profitability.

Next Research Outputs

The first stylised replication is published as The Discrepancy Is Not the Strategy. Once the Ford dataset passes the acceptance gate, the same project can support a real-data Ford fallen-angel event study, a study of apparent versus executable stress basis, and a CDX roll/removal analysis. The mechanics-first CDX study is now published as When One Name Breaks the Basket, with its own reproducible project under research/cdx-index-skew/. The next CDX milestone is to replace the stylised panel with one versioned historical series and reconcile a real auction. Each article should link to an immutable result manifest with the input hashes, configuration, code version, and output tables used in publication.

Related Work