Crypto Carry Research - an honest hunt for a real edge

I ran ~40 methods across 3 waves of crypto alpha research - from WorldQuant-style formula mining to RMT, path signatures, TDA, Hawkes processes and order-flow imbalance - under deflated-Sharpe statistics + cross-engine adversarial verification, proved that the cross-sectional formula alphas are non-stationary and untradeable (in-sample-vs-out-of-sample quintile-shape correlation -0.507, negative gross of cost), and isolated the one structurally defensible edge: convergence-anchored quarterly-futures cash-and-carry basis (every settled contract in-sample positive, ~3%/yr unlevered after costs) - packaged as a read-only, order-gated live engine.

⚠️ Status: a promising research artifact, NOT a proven deployable edge. The surviving edge has a convergence anchor (settlement), but the deployed rule exits 5 days before delivery, so it does not capture the settlement print - it carries residual terminal-basis / roll / execution risk. Headline numbers are out-of-sample and net of modeled costs, but there is no live fill data.

This project's value is not a magic edge - it's the machinery to tell a real edge from a fitted one, used to kill my own hype. Most quant portfolios show one cherry-picked backtest. This one shows the kills - and revised its own headline down when an adversarial review caught its overclaims (see Limitations).

Left: ~40 methods -> almost all killed (formula alphas gross-of-cost = 0; signal inverts out-of-sample) -> 1 surviving structural premium. Right: the basis-carry book, OOS region shaded, OOS Sharpe ~2.9 after basis-leg costs.

---

What survived, and what died (honestly)

Family	Methods tried	Verdict	Why
Quarterly basis cash-and-carry	calendar/term-structure	✅ DEFENSIBLE	settlement gives a convergence anchor -> entry basis is largely earned; every settled BTC/ETH contract in-sample was positive. ⚠️ deployed rule exits 5d pre-delivery -> residual terminal-basis/roll risk (it is not "regardless of price path")
Funding-rate carry	static / hysteresis / x-sectional	⚠️ marginal	real risk premium but regime-dependent (bled ~-1.2 Sharpe in recent OOS); the hysteresis short-spot leg needs borrow (~3.5%/yr) not modeled in the indicator -> not a clean deployable edge
Cross-sectional formula alphas	4,120 formulas, 11 angles	❌ DEAD	non-stationary: IS-vs-OOS quintile-shape corr -0.507, gross-of-cost ~= 0; both textbook cures (stability-selection, walk-forward) also ~= 0
Exotic math	RMT+OU, path signatures, TDA, Hawkes, fracdiff, wavelet, entropy, optimal transport, MFDFA, EVT	❌ DEAD	correct implementations (signature == Chen to 1e-9; MFDFA validated on cascades) - but no extractable predictive content
Momentum / reversal / cointegration pairs	TS/XS momentum, OU pairs, Kalman	❌ DEAD	fragile / huge drawdowns / spreads break OOS
Order-flow imbalance (OFI)	taker + maker market-making	❌ DEAD (retail)	fee-walled as taker; OFI adds zero value to market-making; viability is pure maker-rebate (infrastructure, not signal)

The through-line: within the universe, frequencies, cost models, data sources and method families tested here, no tradeable predictive alpha survived - only a structural risk premium with a settlement convergence anchor. (This is a finding about this search, not a proof that "all public-kline formulas are exhausted" - the program-level multiple-testing penalty is larger than any single PSR suggests.)

---

The credibility backbone (look here first)

1. Deflation is real, not decoration. engine/backtest.py implements the Probabilistic Sharpe Ratio and a Bailey-Lopez-de-Prado Deflated Sharpe benchmark, with a self-test that PASSES a genuine small edge and REJECTS pure noise. Positions are shifted 1 bar (no look-ahead); turnover cost is charged.
2. The non-stationarity kill is the intellectual centerpiece. experiments/alpha_stable.py shows the formula signal is real in-sample (IS-A vs IS-B cross-formula IC corr 0.901 - not noise) yet the tradable quintile shape regime-flips out-of-sample (corr -0.507) and loses gross of cost. Both correct cures were applied and shown to fail: stability- selection (gross -0.33) and monthly walk-forward re-fit (gross ~= +0.08).
3. Cross-engine adversarial verification. Every positive candidate was attacked by an independent Codex refuter (reports/codex_refute_*.md). It killed results I then conceded - e.g. short-spot borrow (~3.5%/yr) wiping a funding-carry "edge"; a COIN-M settlement-print bar being 54% of a candidate's PnL.
4. Deterministic engineering gate (Probity). probity_gate/ is a zero-LLM, mutation-tested ruler gating 6 governance invariants - cost-on-turnover, live-order hard-gating, chronological OOS, deflation primitives, keyless public data, non-normality-robust statistics. The mutation sieve kills all 6 mutants; verdict = PASS (reports/PROBITY_GATE_REPORT.md).

---

The defensible edge: quarterly cash-and-carry basis

• Mechanism: long spot / short dated quarterly future. The exchange settlement rule pulls the future toward the spot index at delivery, giving the entry basis a convergence anchor. This is a risk premium, not arbitrage and not a prediction.
• Honest caveat (corrected after review): the deployed playbook exits ~5 days before delivery (exit_dte=5) to avoid settlement-print noise - so it does not realize the guaranteed settlement convergence. "Regardless of price path / structurally locked" was an overclaim; the exit leaves terminal-basis, roll-timing, liquidity and execution risk.
• Evidence: every settled BTC/ETH quarterly in-sample was positive (the 1-2 nearest contracts are still open/live, shown for completeness, not as settled proof). End-to-end backtest (basis core + funding satellite, basis-leg costs now charged, chronological 60/40 OOS): OOS Sharpe ~2.9, +3.2%/yr unlevered, +9.5%/yr at 3x, over 2024-06...2026-06. (The pre-review headline of Sharpe ~4 / +4.1%/yr omitted basis-leg transaction costs; it has been corrected down.)
• Live: engine/basis_carry_live.py emits a target book (read-only; order placement hard-gated). Cross-exchange scanner engine/xexch_monitor.py covers Binance + OKX + Bitget. Playbook: reports/BASIS_CARRY_PLAYBOOK.md.

---

Limitations (printing these is the portfolio value)

• Exits before settlement. The convergence guarantee only holds if you hold to delivery; the deployed rule exits 5d early, so the edge is convergence-anchored, not settlement- locked. Residual roll/terminal-basis risk is real.
• No live fills. Read-only by design (place_order raises NotImplementedError). Slippage, partial fills, borrow availability, and funding-timestamp execution are modeled, never observed. Nothing here has traded.
• The < 1% max drawdown is cosmetic. It is close-to-close and delta-neutral; it does not capture the intra-bar short-futures-leg liquidation path at leverage. The real risk is margin management. This is a thin (~3%/yr unlevered), capital-intensive carry.
• One OOS regime (~2 years). The basis edge has not seen a violent backwardation / deleveraging cascade out-of-sample. The combined book's period-sensitivity is real (it collapses toward ~0 Sharpe in some sub-windows).
• Program-level deflation is not airtight. Each script deflates within its local variant grid; across 40+ methods and 4,120+ formulas the family-wide penalty is larger than any single PSR = 1.0 suggests. The surviving edge is defended mechanistically, not purely statistically.
• Costs were initially incomplete. The first release omitted basis-leg entry/exit/roll costs in the end-to-end book; an adversarial review caught it and the headline was revised down (Sharpe 3.96 -> 2.94). Cross-exchange / COIN-M inverse legs are scanned live but not yet backtested.

---

Quickstart

pip install -r requirements.txt

# engine self-checks (offline, deterministic)
python engine/stats.py
python engine/backtest.py                # PSR/Deflated-Sharpe: PASSES a real edge, REJECTS noise

# the defensible edge
python experiments/basis_carry_spec.py   # per-contract validation (uses cached data)
python experiments/basis_carry_backtest.py   # end-to-end OOS book, basis-leg costs charged
python engine/basis_carry_live.py        # live target book (read-only; needs internet)
python engine/xexch_monitor.py           # Binance+OKX+Bitget carry dashboard

# the kills (reproduce the honesty)
python experiments/alpha_stable.py       # the -0.507 non-stationarity proof
python experiments/maker_ofi.py          # OFI adds no value to market-making

python experiments/make_showcase.py      # regenerate reports/showcase.png

All market data is fetched from public, keyless, read-only exchange endpoints and cached locally (data/cache/, git-ignored). No API keys, ever.

---

Repository map

engine/        data fetchers (Binance/Bybit/OKX public), math toolkit, backtester
               (PSR/Deflated-Sharpe), live basis-carry engine, cross-exchange scanners
experiments/   the 3 waves of candidates (battery, exotic math, 4120-formula mining,
               OFI) + the surviving carry books + the showcase generator
probity_gate/  zero-LLM, mutation-tested governance gate (6 invariants, verdict PASS)
reports/       REPORT.md (full 23-section write-up, zh), BASIS_CARRY_PLAYBOOK.md,
               PROBITY_GATE_REPORT.md, showcase.png, and per-candidate result JSONs

• Deep dive: reports/REPORT.md - the full research log (Chinese).
• License: MIT (LICENSE). Disclaimer: DISCLAIMER.md - research only, not financial advice, no live trading.