Systematic Corporate Credit · Strategy White Paper

GRANITE

Dislocation-Reversion in Individual Corporate Bonds — buying forced-seller mispricings on the TRACE tape, validated across 22 years
Asset class: US corporate bonds (IG + HY), CUSIP-level Data: TRACE via Open Source Bond Asset Pricing Backtest: Benchmark: LQD
For institutional discussion only. Research prototype and backtest on historical public trade data; not investment advice, an offer, or a solicitation. Simulated past performance does not guarantee future results. Figures are net of the modeled bid–ask spread but gross of fees, financing and taxes. Biases and limitations are addressed explicitly in §8–9.
1 · Executive summary 2 · Opportunity & microstructure 3 · Data & survivorship 4 · Methodology — the honesty harness 5 · The signal 6 · Evidence 7 · Performance vs LQD 8 · Selectivity — a focused book 8.5 · Refinements & honest risk 9 · Bias & overfitting audit 10 · Robust levers & limitations 11 · Reproduction

1Executive summary

The same dealer-market inefficiency that rewards patient buyers of dislocated municipal bonds is present — and independently harvestable — in individual corporate bonds. Buying a bond when it prints several points below its own recent level (a forced-seller dislocation) and holding to reversion earns a statistically significant excess return over random entry in the same bonds, across a full 22-year sample that spans the 2008 GFC and the 2020 COVID shock.

What a manager is buying bias-audited

A repeatable source of +2.1% per-trade excess over a matched control (p<0.001) in a $10-trillion market, validated on the full survivorship-complete universe (every bond, defaulters included), with parameters transferred verbatim from the muni strategy — zero corporate-specific fitting. Attempts to "improve" it with regime/credit overlays were rejected for failing out-of-sample.

2Opportunity & microstructure

US corporates are a ~$10-trillion, over-the-counter dealer market of hundreds of thousands of CUSIPs. Like munis — but with wider spreads and richer credit dispersion — most bonds trade intermittently and customers cross a bid–ask spread on every trade. Every transaction is reported to FINRA (TRACE). Short-horizon timing loses to the spread; a patient buyer of idiosyncratic dislocations is paid a reversion premium for absorbing forced selling (fund redemptions, downgrades, tax-loss and index-exclusion selling).

Fig 1 — Mean customer round-trip return (buy ask → sell bid) by holding horizon, unconditional across all bond-days. Break-even ~4–6 months; the strategy holds ~1 year.

3Data & survivorship

Source: Open Source Bond Asset Pricing (Dickerson et al.) — a free, processed daily corporate-bond panel built from TRACE, with clean price, bid, ask, yield and credit spread. The bid/ask lets us model execution faithfully (buy at the ask, sell at the bid).

Survivorship-complete

The panel is the full TRACE tape, including bonds that default, get called or go quiet. 70% of bonds stop trading before the sample end and are retained with their final prices; 8% ever crash below 50; and 7–14% of strategy positions exit "stale" — the bond stopped trading mid-hold — taking the realized loss. The edge is measured after those losses.

4Methodology — the honesty harness

5The signal

GRANITE entry rule

Enter when a bond's ask prints ≥3 points below its trailing 60-day median clean price and it clears the liquidity gate. Fill at the ask. Exit at the bid after a ~1-year hold. The rule is the muni KEYSTONE specification applied unchanged — a deliberate out-of-sample transfer test across asset classes.

6Evidence

6.1 Signal strength is monotone in depth

Fig 2 — Excess vs matched control by entry threshold (full sample).
Fig 3 — Excess vs control by era; positive or flat in every regime.
Entry thresholdTradesWinMean/tradeExcessp
Full sample 2002–2025. Excess = strategy − matched random-entry control in the same bonds.

6.2 In-sample → out-of-sample

WindowTradesWinMean/tradeExcess vs ctrlp

6.3 Regime robustness

EraTradesWinMean/tradeExcessp

7Performance vs LQD

Equal-weight portfolio of open positions (≥3-point operating point) vs LQD total return over the same window.

Fig 4 — Growth of $1, total return. See §9 on the path-smoothing caveat.
TotalCAGRMax drawdown*
*Chart drawdowns use linear intra-trade attribution and are understated: re-marking daily at actual mid prints gives maxDD −32.7% (full book) / −31.3% (≤5y), monthly Sharpe vs T-bill 0.41 / 0.46. Total return & CAGR are unaffected (realized from bid/ask fills). See §8.5.

8Selectivity — a focused short-duration book

The full-universe book carries ~3,100 concurrent positions — more than a real desk would run. Can a much smaller, higher-conviction book keep, or beat, the edge? We tested point-in-time filters on credit quality (entry-day credit spread) and duration (years to maturity), each baked into the eligibility gate so the matched control trades the same selected subset. The answer is decisive — and it overturns the intuitive guess that "trade only the safest bonds."

Credit quality — filtering to high grade backfires

Entry credit spreadOOS tradesWinMean/tradeExcess

Top-IG dislocations have no out-of-sample timing edge and even lost money outright (the 2022 rate selloff crushed long high-grade). The reversion alpha lives in crossover/HY names; only the very deepest distress (spread >5%) is worth trimming, and only as a tail-risk knob.

Duration — the edge concentrates in short bonds

Years to maturityOOS tradesWinMean/tradeExcess

Short-dated (≤5y) dislocations must pull to par, so they revert — OOS excess +3.5%, double the full book, on ~30% of the trades. Long bonds (>12y) show zero OOS edge: their price is driven by rates, not idiosyncratic reversion.

The focused operating point (≤5y) vs the full book

Short-dated selection is not an overfit overlay: it is monotone across the maturity spectrum, economically structural (pull-to-par), improves both IS and OOS, and — unlike the rejected regime gate in §9 — repairs the 2008 GFC out-of-sample rather than curve-fitting it. Fewer names, higher conviction, and the base book's one weak regime is gone.

Fig 5 — Growth of $1: focused ≤5y book vs LQD, total return.
BookAvg pos.TotalCAGRMaxDD*OOS excess
Era (≤5y book)TradesWinExcessp

*Drawdown understated by linear intra-trade attribution — honest daily-mark figures in §8.5; total & CAGR realized from bid/ask fills.

8.5Refinements & the honest risk picture (second research round)

A second research program — six agent "trading lenses" proposing 24 specs, 13 strategy families screened in-sample with pre-registered kill gates, survivors given exactly one out-of-sample look — searched for a higher-Sharpe construction and stress-tested this strategy's accounting. Three results matter:

1 · Honest mark-to-market risk (correction)

The equity-curve drawdowns above use linear intra-trade attribution, which smooths the path. Re-marking every position daily at its actual mid prints (stale marks flat, monthly Sharpe vs T-bill):

bookchart maxDDhonest maxDDSharpe
full ≥3pt−14.1%−32.7%0.41
focused ≤5y−15.4%−31.3%0.46
≤5y + issuer cap—−30.0%0.58

Total return and CAGR are unaffected — they were always realized from bid/ask fills. Real risk is ~2× the smoothed path; risk-adjusted return is ordinary, not exceptional.

2 · GRANITE-CL — the validated refinement

On the issuer-capped ≤5y book, accept the fill only if the ask ≤ latest prior mid + 0.25 — never chase a bounced price (the control faces the same fill cap). One frozen parameter; the cap is an implicit depth filter, and depth was monotone from the first sweep.

windownmean/trexcessCAGRSharpe
IS 2003–152,018+12.62%+4.56%+11.45%0.99
OOS 2016–25 (one-shot)2,733+10.98%+5.48%+8.05%0.56
full 2003–254,592+10.54%+4.58%+8.88%0.73

OOS excess above IS — no decay. All p<0.001. The rule also transfers in reverse to munis (excess IS +1.97→+3.36%, OOS +2.99→+3.99%) — four independent validating samples. Diagnostics: cap monotone, excess positive in all 6 eras (incl. GFC +4.66% and 2021–23 +1.37%), strongest in the most liquid quartile.

GRANITE-XL — depth weighting + recovery exit

Same entries, two more frozen rules: size ∝ dislocation depth (capped 0.5–2×) and sell at the first bid ≥21d in once the day's mid recovers to the entry-day median (455d stop). Third one-shot OOS batch, paired on identical entries:

GRANITE-XLmean/trholdCAGRSharpemaxDD
IS 2003–15+6.34%271d+19.24%1.15−41.6%
OOS 2016–25 (one-shot)+6.87%266d+17.13%1.03−34.3%

Faster recycling: deep entries revert early, the exit harvests and redeploys — per-trade halves, compounding doubles. (The recovery-exit idea had one prior OOS glimpse in another configuration — disclosed; all rules frozen before the batch.)

2026-08 audit: those rows accrue carry at the panel's median-yield proxy; re-priced on recovered real coupons the same fills give IS +17.8%/1.09, OOS +12.6–14.2%/0.84–0.86 (windowing convention), full +14.7%/0.93. Implementable live-protocol replays land at OOS +11–14% CAGR — budget those, not the headline. See XL_AUDIT.md.

3 · The negative result, documented

The search's target — Sharpe ≈3 / CAGR ≥10% — was not met, with the mechanism demonstrated rather than asserted: the bid-ask toll is paid twice per round trip (it killed every short-horizon sleeve, including one that was significant and monotone in-sample yet still died out-of-sample); all surviving sleeves are long the same credit-beta factor (OOS correlations 0.69–0.85, so combination cannot lift Sharpe); vol-targeting trades CAGR for Sharpe without leverage. Clean kills: fire-sale reversal, spread-widening reversion, carry ladders, issuer relative-value, momentum, fallen-angel proxy, new-issue concession, index-exclusion and hold-to-redemption constructions, vol-managed sizing, calendar effects. The refinement path lifted the frontier to Sharpe ~1.0–1.2 at CAGR ~17% OOS (GRANITE-XL) — the CAGR-10 target is met out-of-sample; Sharpe 3 is not, because every position remains long the single credit-beta factor through a −34% crisis drawdown, and tripling Sharpe requires hedging instruments (CDS, futures, repo leverage) this dataset does not contain. Full write-up: corps/NEW_STRATEGY_SEARCH.md.

9Bias & overfitting audit

The checks a credit-quant desk would demand. Full detail in corps/CORP_AUDIT.md.

RiskStatusHow addressed
Survivorshipcontrolledfull TRACE tape; defaulters retained; 7–14% stale-exit losses taken
Selectioncontrolledall 55,545 bonds, not a top-N cut; removing our earlier cut raised the excess
Look-aheadcontrolledlagged median, trailing gate, next-day entry, past-only exits
Overfitting (params)controlledparameters transferred verbatim from munis — zero corporate fitting
Overfitting (signals)demonstratedregime/credit overlays rejected for failing OOS (below)
Transaction costsmodeledactual daily bid/ask fills

Rejected "improvements" — the anti-overfitting test

Two economically-motivated overlays were tested and rejected: the market-regime gate looked strong in-sample and "fixed" the 2008 GFC but degraded out-of-sample (the textbook overfitting signature); the per-bond credit filter hurt outright. We publish the base.

VariantIS excessOOS excessVerdict

10Robust levers & limitations

Robust operating levers (OOS-validated)

Deeper thresholds concentrate the alpha (fewer, higher-conviction trades); a dynamic recovery-exit cuts average holding ~35% at similar annualized return.

Limitations (disclosed)

  • Carry proxied by median yield (nets out of the control excess).
  • Equity-curve drawdown smoothed by linear attribution; returns are realized.
  • Capacity: full-universe number includes illiquid names; a live book tiers by liquidity.
  • Prices are OSBAP cleaned daily VWAP + bid/ask, not raw ticks.
  • Systemic crises (GFC) are the full book's weakest regime; the focused ≤5y book (§8) repairs it out-of-sample, leaving the 2021–23 rate selloff as the soft spot.

11Reproduction

# free data (no WRDS): the daily TRACE panel is committed to the repo, year-partitioned python corps/scripts/build_osbap_panel.py # or use corps/data/panel/*.parquet python corps/research/osbap_full.py # bias-free full-universe backtest python corps/research/osbap_improve.py # overlay tests (rejected) python corps/research/osbap_improve2.py # robust levers (depth/hold/dynamic exit) python corps/research/finalize.py # equity curves + this page's data

GRANITE — corporate-bond dislocation-reversion. Data: Open Source Bond Asset Pricing (TRACE). Full method, audit, and per-era detail in corps/CORP_FINDINGS.md and corps/CORP_AUDIT.md. Generated from docs/corps_data.json; every figure recomputes from the panel.

For institutional discussion only · not investment advice · simulated past performance is not indicative of future results.