Volatility-Managed Long-Only Majors Basket (BINANCE USD-M, 6 Majors Equal-Weight, Scale Gross Exposure by Realized-Vol Regime — Full in Calm / Half in Elevated / Cash in High-Vol — Weekly, Pure OHLCV, 2-Parameter)

A LONG-ONLY, weekly-managed equal-weight basket of 6 data-verified BINANCE USD-M majors (BTCUSDT, ETHUSDT, SOLUSDT, BNBUSDT, AVAXUSDT, LINKUSDT) whose distinguishing mechanism is a VOLATILITY-MANAGED gross-exposure overlay (Moreira & Muir, 'Volatility-Managed Portfolios', JF 2017) — a documented, mechanically-simple Sharpe-raising technique ORTHOGONAL to the momentum and carry factors already deployed. Instead of predicting direction, it scales gross exposure DOWN as recent realized volatility rises (volatility is persistent and reliably precedes drawdowns, especially in crypto) and holds full exposure when markets are calm. This is a re-attempt of a previously abandoned hypothesis whose abandon was an ENGINE BUG (corrupted close-date daily-return Sharpe, fixed 2026-07-24 commit 59c8a73 — Sharpe/vol now from the per-bar MTM equity curve), not a strategy verdict; QA credited the code. It is engineered against this session's hard lessons: PURE OHLCV only (best-surviving tag; no OI/liquidation/funding data walls that abandoned the liquidation and quarterly-basis siblings), MULTI-ASSET equal-weight (escapes the 0/213 single-symbol no-edge class), STANDARD BINANCE USD-M linear venue at leverage 1.0 max (avoids COIN-M venue-mismatch and tick-data implementation failures), and a plain hold-and-resize structure with only 2 tunable parameters (vol lookback, regime cutoffs) to resist overfit (the 202x killer). Discrete 3-state regime switching keeps turnover — and fees — low. Value proposition is a higher Sharpe and lower max-drawdown than an unmanaged equal-weight majors basket, NOT raw return; the overlay's payoff (cutting exposure into persistent high-vol drawdown regimes such as 2022) is distributed across every volatile bear stretch, which is the direct rebuttal to the single-regime-2021 artifact that killed the market-neutral-momentum sibling.

Implements the hypothesis exactly: a Moreira-Muir volatility-managed overlay on an unmanaged equal-weight majors basket, pure OHLCV, multi-asset, standard BINANCE USD-M linear venue at leverage 1.0, weekly cadence, discrete 3-state switching for low turnover. Key implementation choices against the session's failure modes: (a) the vol threshold is a rolling PERCENTILE of the strategy's own vol history, not an absolute constant, so the regime split is scale-invariant across the 2020-2026 vol level shift (L5); (b) there is NO compound entry conjunction - the book is long the full basket unless vol is in its top quintile, so the sandbox window will produce many trades rather than looping on zero (L25/L21); (c) warm-up is only 22 bars for the signal and ~112 daily bars before the first trade, a small fraction of the 2400-day sandbox window; (d) all per-bar work is O(1) on bounded deques (no recomputation over accumulated history), so the 300s smoke limit is not at risk; (e) sizing is target-notional relative to get_account_equity() with gross capped at 1x and long-only, so the compounding-blowup pattern of L15 cannot occur, and SELL orders can only reduce an existing long; (f) the rebalance schedule is derived from bar.ts_init (epoch-day % 7), never from a process-relative bar counter, so it is restart- and replay-identical. All legs are currently-listed liquid Binance USD-M perps with full daily history from 2020, avoiding the missing-leg data wall. Venue note: the strategy is long-only at leverage 1.0, which by the default routing rule would suggest BINANCE_SPOT, but the hypothesis explicitly specifies BINANCE USD-M and the futures venue is also the cheaper one here (0.10% round-trip vs 0.20% spot), which materially helps the per-trade fee floor on a rebalancing basket - so I kept the declared venue. Only three parameters really drive the edge (vol_lookback, calm_pct, high_pct); the rest are sizing/cadence hygiene, and every hard clamp in the code is declared in _param_bounds.

Long-only vol-managed majors basket (L45/L49/L36 family, 0/84 survivors) that is both mis-sized and falsified against its own benchmark — not worth 2 hours of optimization. The sizing is broken: exposure_pct is 478% despite the stated 1.0x gross cap, so the book runs ~4.8x leverage and posts an impossible +2067% single day on 2023-11-16 (annual_returns 2023 = 2067.87; return_kurtosis 117.5) — a compounding/leverage artifact, not edge. That drives a 77.6% max drawdown (CI to 90.8%, a hard-abandon level per L19), 2022 -97% and 2026 -53%. Critically, the strategy's explicit value proposition — higher Sharpe / lower drawdown than an unmanaged equal-weight basket — is falsified: information_ratio -0.817 and alpha -0.051 versus its equal-weight-basket benchmark, so the vol overlay makes things WORSE than doing nothing. There is no edge (profit_factor 0.978, negative expectancy, Sharpe 0.436 with CI straddling zero) and the +131% headline is >100% open-position unrealized (end_unrealized_pct 141). Even fixing the exposure bug returns a beta-carried long-only crypto basket that loses to its own benchmark. Failure pattern: risk_reject/no_edge long-only cross-sectional basket, leverage/compounding artifact + negative IR vs benchmark (L45/L49/L36/L19).

Verification failed (Layer 4 — QA review) [class=premise_falsified]: - The hypothesis's entire value proposition is 'higher Sharpe and lower max-drawdown than an unmanaged equal-weight majors basket, NOT raw return.' The sandbox benchmarks against exactly that (benchmark_type=equal_weight_basket, benchmark_meaningful=true) and REFUTES both legs: alpha -0.0217 and information_ratio -0.66 mean the overlay delivers LOWER risk-adjusted return than the unmanaged basket, and max_drawdown is 77.6% (no protection — the basket's own 2022 DD is ~75-80%). Mechanistic cause: avg_position_pct ~15.6% ≈ full exposure (1.0x/6 legs = 16.7% per leg) — the overlay barely de-risks. This is structural to the specified percentile-rank design: ranking realized vol against its own trailing 365-day distribution means that in a persistent high-vol bear (2022) the distribution recalibrates, the percentile falls back below high_pct, and the book re-risks into the ongoing drawdown it was built to avoid. The promised vol-managed effect is absent and net-negative vs the benchmark. Faithful implementation of the hypothesis design; the design itself does not produce its effect. - Title claims '2-Parameter' but the edge uses at least three tunables (vol_lookback, calm_pct, high_pct) and _param_bounds declares eight. Minor scope drift; relevant to overfit risk if this class is re-attempted.