Strategies · GRANITE white paper · GRANITE-XL · KEYSTONE (munis)
SYSTEMATIC CORPORATE CREDIT · FINAL OPERATING POINT

GRANITE-XL

Dislocation-reversion in individual U.S. corporate bonds, refined through four research rounds, three disclosed out-of-sample batches, ~19 killed strategy families, and a bidirectional muni↔corporate transfer test. Every number below is at honest daily marks (positions re-marked at actual mid prints, monthly Sharpe vs T-bill) with fills at the actual customer ask (buy) and bid (sell).

The five rules

Dislocation entry. Buy when a bond's customer-ask prints ≥3 points below its trailing 60-day median clean price and it passes a trailing-liquidity gate (≥8 active days / 90). Parameters transferred verbatim from the muni strategy — zero corporate fitting.
Short duration only (≤5y). Short bonds must pull to par, so dislocations revert; long bonds are rates bets. OOS-validated; repairs the 2008 GFC.
Limit entry — never chase. Accept the fill only if the ask ≤ the latest prior mid + 0.25. An implicit depth filter (accepted fills average 4.3 pts deep); monotone in the cap; transfers in reverse to munis.
One position per issuer; size ∝ depth (weight = depth/3, capped 0.5–2×). Issuer diversification beats raw breadth.
Recovery exit. Sell at the first bid ≥21 days in once the day's mid recovers to the entry-day median; 455-day hard stop. Harvest the reversion, recycle the capital (holds ~380→266 days).

Performance — honest mark-to-market

Growth of $1, GRANITE-XL (depth-weighted, honest daily marks at real mid prints, carry at recovered real coupons — 2026-08 audit) vs LQD (IG corporates — the like-for-like credit benchmark) and AGG (US aggregate bond market). AGG launched 2003-09, nine months after the backtest starts, so its line begins there rebased to $1 rather than being backfilled; the table below measures all three from that date. Log scale — equal vertical distance is equal percentage change, without which the two benchmarks collapse into one flat line. The steep final leg (2024–25) runs on a wound-down book of a few positions and should not be read as strategy performance (see Year by year).

Strategy vs both bond benchmarks — common window from 2003-10

SeriesCAGRMax DDGrowth of $1
Apples-to-apples: all three measured over identical dates, starting at AGG's inception. LQD (investment-grade corporates) is the like-for-like credit benchmark — same asset class, same spread risk. AGG (US aggregate: Treasuries + agency MBS + IG credit) is the broad bond-market benchmark, lower yielding and lower drawdown. The strategy beats both on return; its drawdown is deeper than either.

Per-window detail

WindowTradesMean/tradeWinHoldCAGRSharpe (m)MaxDD
Swipe the table sideways on a narrow screen to see all columns.
Validation protocol. The entry family was designed on munis and 2003–2015 corporates only; every refinement was admitted to out-of-sample (2016–2025) through a pre-registered in-sample gate and tested once. OOS excess vs a matched random-entry control facing the same fill cap: +5.48% per trade, p<0.001 — above the in-sample figure (no decay). Disclosure: the recovery-exit idea had one earlier OOS glimpse in a different configuration; era/ML decompositions of already-spent OOS windows are reporting, not selection.

BEDROCK-V — the validated upgrade (2026-08)

BEDROCK-V = GRANITE-XL entries + two gates, found by a literature sweep and validated under this repo's pre-registered protocol: (G1) spread-value — take the dislocation only if the bond's credit spread is at or above its same-day maturity-bucket median (cheap vs peers; the one signal family that survives the corporate-bond replication-crisis literature), and (G4) issuer-curve depth — the bond must be dislocated ≥2 pts deeper than the median of its own issuer's sibling bonds (bond-specific forced selling, not issuer-wide news). Same engine, same honesty rules, real coupons, lagged recovery exits. Fewer, better trades.

BookTradesMean/tradeWinCAGRSharpe (m)MaxDD
Validation trail (one-shot OOS, specs frozen in git before results). Robustness: CAGR 17.8–18.6% / Sharpe 1.04–1.05 across every gate perturbation tested (G1 at the 40th/50th/60th cross-sectional percentile × G4 gap 1/2/3 pts) — no knife-edge. Era decomposition positive in all six eras; best vintages are the crises (2008–09 +13.9%, 2020 +12.9% per trade). Killed along the way by their own pre-registered tests: the 4-gate stack, spread-momentum and volume-stress gates, an index-exclusion sleeve, a muni de-minimis overlay, calendar overlays, issuance-pressure. Full record: BEDROCK_RESEARCH.md.

What to expect live — replay × slippage planning rows

Protocol (OOS 2016–24)TradesMean/tradeCAGRSharpe (m)MaxDD
Chronological live-protocol replay (issuer capacity checked against actual open positions, limit filter before capacity, gates at signal time, real coupons) with entry/exit haircuts h. Plan on the h=⅛ row: ≈+12.5% CAGR / Sharpe(m) ~0.7 — not the pipeline headline. GRANITE-XL's own replays under identical accounting: +10.7–11.5% / 0.59–0.74. The replay book is ~2× the pipeline (5,983 fills full-window) — the gates concentrate quality without shrinking live capacity.
2026-08 audit — plan on the corrected rows. The headline rows accrue carry at each bond's median yield (the panel omits coupons); the italic real coupons rows re-price the identical fills with recovered per-bond coupons (validated inversion — near-par error 0.05 pt, 99.95% premium-bond consistency). Monthly Sharpe is smoothed by stale marks (monthly autocorr 0.38); annual-frequency Sharpe is ~0.5–0.65. A chronological live-protocol replay (implementable issuer-lockout rules, prior-day recovery trigger, real coupons — corps/research/live_replay_*.py) lands at OOS +11.5%/0.74 (tight ~13-month lockout) to +10.7%/0.59 (capacity recycled at the actual exit): the published pipeline's extra selectivity comes from limit-rejected candidates consuming issuer capacity, and should be treated as model risk. Budget OOS CAGR +11–14%. Full write-up: XL_AUDIT.md.

Regime robustness

XL per-trade returns by era

EraTradesWinMean/trade

Entry-timing excess vs control by era (CL book)

EraExcessp

Positive in all six eras, significant in five — including the 2008 GFC and the 2021–23 rate selloff, the two historically weak regimes for credit mean-reversion.

Why it works — the response-curve atlas

Forward returns (mid + accrual) after each event type, in excess of a matched baseline on the same universe. Three laws: reversion is monotone in depth and slow (the payoff lives at 63–365 days — short-horizon trading dies to the bid-ask toll); volume alone is noise; and momentum is inverted (bonds at their 250-day high underperform — the par ceiling caps upside). The five rules are exactly what this table prescribes.

Condition21d63d126d252d365d

The transfer test — same rules, different market

The strongest anti-overfitting evidence available: the XL stack applied verbatim to municipal bonds (different asset class, issuers, data vendor, backtest engine).

KEYSTONE (munis)TradesWinMean/tradeExcess vs ctrlCAGRMaxDD*
*Muni NAV uses linear intra-trade attribution (the muni engine has no daily mid-mark); corporate figures above are honest daily marks. Muni OOS window (2023–26) was a benign muni tape — drawdowns near zero across all variants.

Anatomy of a trade — historical examples

Real trades from the backtest, with the bond's full surrounding tape. Selection is by percentile of realized return among tape-dense trades (90th / median / 10th), plus the best COVID-2020 entry and a median stale hard-stop loser — representative outcomes, not cherry-picked winners. Gold dots are days customers bought (dealer ask); slate dots are days customers sold (dealer bid); the dashed line is the trailing 60-day median that defines the dislocation.

Quant appendix

Everything under the hood: where each point of return comes from, the shape of the trade distribution, monthly moments, the worst episodes, and what the book actually holds. The decomposition is an exact identity — the four components sum to the per-trade total.

Return decomposition (mean per trade)

ComponentMeanShare

Carry = coupon accrual over the hold. Entry edge = discount captured vs the entry-day mid (the limit cap's contribution). Reversion = mid-to-mid price recovery — the alpha engine. Exit cost = selling at the bid below mid.

Per-trade return distribution

Monthly return moments

Worst drawdown episodes

PeakTroughRecoveredDepth

Book structure & execution

Year by year

YearBook returnTrades enteredMean/trade (entry vintage)
Read the tail with care. Entries stop 455 days before the panel ends, so 2024–25 enter no new trades and the book runs down to a handful of positions (mean ~24 in 2024, ~4 in 2025 vs a ~112 median in normal years). Those last two years' returns — and the steepest final leg of the equity curve — are single-name risk, not strategy performance; 1.9% of all invested days carry fewer than 10 open positions. The IS/OOS and common-window figures above are the ones to plan on.

Execution & live replication

Already inside every backtest number: fills pay the full printed bid–ask (buy at the customer-ask VWAP, sell at the customer-bid VWAP); no fill unless the required side actually printed (no phantom liquidity, 7-day fill window or no trade); entry is the day after the signal (no latency assumption); dead bonds exit at their last real bid. What is not automatic is competition for the print and size — modeled below.

Slippage stress — pay up h to win the print

Haircut hMean/tradeWinCAGRMaxDD

Capacity at 25% participation of printed volume

Position capped at 25% of the entry day's printed volume in that bond. Conservative: a patient buyer can also work orders across the multi-day dislocation window, and reported TRACE volumes are top-coded (large trades shown capped), so true capacity is higher.

Live-desk checklist — what a replication must handle

Honest limits

What this strategy is

  • One real phenomenon — forced-seller dislocations in short-dated credit revert — executed with discipline.
  • OOS CAGR +17.1% with monthly Sharpe ~1.0 at the panel's yield-proxy carry; on recovered real coupons the same fills give OOS +12.6–14.2% / Sharpe(m) 0.84–0.86 (the range is the windowing convention: slicing the full run by entry date vs re-running the OOS window standalone), annual-frequency Sharpe ~0.6 (2026-08 audit — see XL_AUDIT.md). Excess robust in every era and in the most-liquid quartile (capacity-viable).
  • Validated by kill-tested search: ~19 strategy families died on pre-registered gates around it.

What it is not

  • Not market-neutral: long credit beta throughout; honest maxDD ~−34% in systemic crises. Demonstrated irreducible: nine downside-protection mechanisms (floors, stops, sibling filters, ML tail-avoidance, exit loosening, deployment pacing) all failed pre-registered gates — every one either sold the crisis returns or left the marks untouched. Size the book so the drawdown is tolerable; that is the price of the CAGR.
  • Not Sharpe 3: without CDS/futures hedging, shorting, or leverage, Sharpe ~1.0–1.2 is this instrument set's frontier — demonstrated, not assumed.
  • Backtest fills assume the patient trader captures the posted customer ask/bid (OSBAP daily VWAP), not guaranteed executable ticks.
For institutional discussion only — not investment advice. Simulated results on the Open Source Bond Asset Pricing daily TRACE panel (survivorship-complete, 55,545 bonds, 2002–2025); net of modeled bid–ask spread, gross of fees/financing/taxes. Full research trail: corps/NEW_STRATEGY_SEARCH.md, audit in corps/CORP_AUDIT.md, white paper at corps.html. Past performance does not guarantee future results.