# Where the Edge Actually Is: A Walk-Forward, Anti-Mirage Study of Systematic Crypto Strategies (BTC, ETH, SOL)

*Degen Daycare / Hector Research Labs — working draft v2, 2026-07-02 (evolving under DD-001 Research Evolution Loop; gate report: MYTHOS/QA/PAPER_GATE_R1.md). All results from real data; honest by construction.*

## Abstract
We evaluate a family of systematic strategies on BTC, ETH, and SOL using strict walk-forward validation
(180-bar train / 3-bar embargo / 30-bar test, 53 unseen daily folds, ~5y Coinbase data, 12 bps/flip costs).
We find that **single-asset directional prediction has no durable net edge** — every momentum/mean-reversion/
breakout/factor-neutral price model scores ≤ 4/7 on a 7-check readiness gate, failing in near-identical ways
across assets, indicating the limitation is the market, not the model. The durable edge is **delta-neutral
perpetual funding carry**, and it is **asset-specific**: SOL is a trend asset; BTC and ETH are carry assets.
Carry alone clears the strict gate on BTC; ETH requires its **native staking yield** (a long-stETH / short-perp
"LST-basis carry") to clear. We subject the winners to an anti-mirage gate — **Deflated Sharpe Ratio** (Bailey &
López de Prado 2014), **Probability of Backtest Overfitting via CPCV** (Bailey et al. 2017), and cross-asset
generalization — and a stress battery. BTC's carry+overlay champion is robust (DSR 0.99, PBO 1.6%, 9/9 parameter
combos hold, cost cliff ~30 bps); ETH's is real but marginal (DSR N-sensitive). We argue the correct design
philosophy is a transparent **portfolio governor** that allocates to carry and uses directional signals only as a
small regime-gated overlay — keeping the edge simple and the sophistication in risk allocation.

## 1. Data
- Prices: daily OHLC, Coinbase (BTC/ETH/SOL), ~1,800 bars each (~5y).
- Funding: real perpetual funding, Binance USDⓈ-M via the Binance Vision static archive (live API geo-blocked;
  OKX/dYdX used for live). ETH 86% positive days, ~14.7%/yr mean; BTC 86%, ~12.7%; SOL 72%, ~4.8%.
- Staking: real Lido stETH APR history, DefiLlama yields (mean 3.49%/yr, range 2.24–11.77%).
- Orthogonal: Fear&Greed, Solana DEX volume (tested; not load-bearing).

## 2. Method
Walk-forward: pick on TRAIN, grade only on the next unseen TEST window, 3-bar embargo to kill leakage, 53 folds.
Costs: 12 bps per position flip + a 90% funding-capture haircut. Bar-level Sharpe annualized. The **7-check
readiness gate**: avg-fold Sharpe ≥ 0.35 · ≥55% folds profitable · beats buy&hold · MC blow-up ≤ 10% · 5th-pct
MC maxDD ≥ −30% · turnover < 2/day · ≥ 20 folds. The **/10 championship gate** adds: (8) Deflated Sharpe ≥ 0.95,
(9) PBO < 5% via CPCV, (10) cross-asset hold. N (trials) is counted per-asset (the selection is per-asset).
Carry sleeve: long spot / short perp when funding > 0 (flip when < 0), 3-day-MA sizing, asymmetric entry / wide
hold band (low turnover), stress/choppy kill-switches. ETH adds staking yield on the long-spot leg. The champion
adds a 0.15-weight trend overlay gated by a 20-bar efficiency-ratio regime score.

## 3. Results
| Asset | Directional (best) | Carry-alone | Champion (carry+overlay[+staking]) | /10 |
|---|---|---|---|---|
| SOL | **Gated Trend 7/7** (Sharpe 0.91) | 5/7 | trend "champion" — **a demonstrated mirage** | **8/10** (❌ DSR **0.029**, ❌ PBO **79.3%**) |
| BTC | ≤ 3/7 | **7/7** (Sharpe 4.09) | META-ALLOC | **10/10** (DSR 0.996, PBO 1.5%) |
| ETH | ≤ 4/7 | 6/7 (Sharpe 2.74) | **SAM+META** LST-basis (Sharpe 4.61) | **10/10** (DSR 1.00*, PBO 2.4%) |

\* ETH DSR is N-sensitive (0.77 at N=250 → ~1.0 at the correct per-asset N≈40); BTC's DSR is robust to N.
All values re-verified against the 2026-07-01 gauntlet records (`docs/CHAMPION_*.json`).

**The mirage, caught by our own gate (this is the paper's thesis performing itself):** SOL's directional
strategy passes the full 7-check readiness gate — every box green, Sharpe 0.91, +51.6% OOS against a −29.0%
buy-and-hold — and then fails the anti-mirage battery catastrophically: Deflated Sharpe **0.029** against a
0.95 bar, and a **79.3%** probability of backtest overfitting via CPCV. Under a 7-gate regime we would have
funded it. Under the 10-gate regime it is correctly named what it is: forty trials' worth of luck wearing a
championship belt. A base readiness gate is a *necessary filter, not a sufficient verdict* — this single result
is, in our view, the strongest argument in the paper.

**3.1 Multi-agent replication and the shared library (2026-07-01).** We instantiated two further autonomous
agents with distinct strategy personalities — "Sam" (aggressive breakout/momentum) and "Lianna" (patient deep
mean-reversion) — and ran them through the identical gauntlet. The directional-mirage pattern *replicated*:
Sam's momentum champion also passed 7/7 (Sharpe 0.80) and also failed anti-mirage (DSR 0.354, PBO 39.2%);
Lianna's best solo configuration managed only 4/7 (Sharpe −0.57) on a trending asset hostile to her style.
We then gave every agent league-access to a shared **house library** of the ecosystem's gated strategies
(the carry sleeve and hybrids). In a single run, Lianna's champion configuration became a house-hybrid
("carry when funding pays, else her own champion") scoring 6/7 with Sharpe 0.95 — the weakest agent lifted
by *knowledge transfer rather than retraining*, while retaining her own style brains. We propose this
**competition + shared-library** structure as a reusable framework: agents compete on identical unseen data
(which exposes mirages), while proven, gate-cleared strategies propagate to all agents as infrastructure
(which compounds honest edge). Records: `CHAMPION_SAM_DAILY.json`, `CHAMPION_LIANNA_DAILY.json`.

**Round-by-round honesty (the League):** independent LLM "desks" (GPT/Grok/Gemini/Claude/Kimi/Copilot) each
proposed strategies; we implemented and scored them identically. The most-confident Round-1 idea (a BTC-residual
factor-neutral momentum, two desks' #1) was the WORST of 16 brains; a "maybe"-rated volatility-regime sizing was
the only directional idea that helped. Confidence did not predict edge.

## 4. Robustness (stress battery)
- **Friction:** BTC champion survives 2× cost (24 bps, 6/7) and 30%-thinner funding (still positive); **cost
  cliff ≈ 30 bps/flip** (2.5× margin over modeled 12). ETH cliff ≈ 24 bps.
- **Parameters:** BTC **9/9** sweep combos hold 7/7 (plateau, not knife-edge); ETH 6/9.
- **Negative-funding regime:** inverting funding, the delta-neutral sleeve flips to short-carry and still earns
  (BTC Sharpe 7.6, 0% blow-up) — robust to funding sign.
- **Sub-period:** edge is regime-dependent (negative in 1 of 3 sub-periods for BTC — a compressed-funding stretch).

## 5. Limitations (honest)
Backtest models funding income + flips + capture haircut but NOT per-venue fees, slippage, basis-at-entry/exit,
LST depeg, or smart-contract risk. avg-fold Sharpe is inflated by short-window annualization (we report overall
OOS Sharpe as the trustworthy figure). The /10 gate is the strongest in-sample proxy; the definitive test is a
live forward (out-of-sample-in-time), which we run separately. Nothing here is funded; this is not financial advice.

## 6. Conclusion
Stop trying to predict crypto direction; harvest the structural carry premium and govern it. The edge is simple,
asset-specific, and exogenous (funding + staking); the value is in transparent risk allocation, not ever-more-complex
prediction. BTC carry is the robust champion; ETH carry needs its native staking yield to clear and is marginal.

## Reproduce
`node training-camp/funding-fetch.mjs ETH|BTC|SOL` · `node training-camp/staking-fetch.mjs` ·
`node training-camp/gauntlet.mjs daily BTC --stress` → writes `docs/CHAMPION_*_DAILY.json` (readiness.gate10/.league/.stress).
Dashboard: `arena/research.html`. Desk submissions & round logs: `docs/LEAGUE_*.md`, `docs/RESEARCH_DESK_*.md`.

## References
Bailey & López de Prado (2014), *The Deflated Sharpe Ratio*. Bailey, Borwein, López de Prado & Zhu (2017), *The
Probability of Backtest Overfitting*. Blitz, Huij & Martens (2011), *Residual Momentum*. Schmeling, Schrimpf &
Todorov (2024), *Crypto Carry* (BIS WP 1087). He & Manela (2022), *Fundamentals of Perpetual Futures*.