EthSpotVolatilityManagedRiskOnOffLongFlat

ETH Spot Volatility-Managed Risk-On/Off, Long-Flat (Single-Instrument ETHUSDT.BINANCE_SPOT, Daily Bars, Hold ETH Only in Low-Realized-Vol Uptrend Regimes and Sit in Cash When Volatility Spikes, Vol-Targeted Sizing, Pure OHLCV, 3-Parameter)

A LONG-FLAT, single-instrument regime-timing strategy on ETHUSDT.BINANCE_SPOT that implements the well-documented volatility-managed-portfolio anomaly (Moreira & Muir): risk-adjusted returns are systematically higher in LOW-volatility regimes and the worst drawdowns cluster in HIGH-volatility regimes, so scaling exposure inversely to realized volatility raises Sharpe and cuts tail risk out-of-sample. It fills the under-represented BINANCE_SPOT venue (12.0% vs >=15% target) using nothing but daily OHLCV — the best-surviving data class (no_supplementary_data, 4%). It is deliberately the OPPOSITE of the buy-the-dip / accumulate-the-flush spot ideas that overfit (it sits OUT during capitulation, not into it) and is NOT a price-trend chaser, a rank-rotation basket, a VWAP/RSI fade, or a supplementary-feed signal — so it avoids every enumerated failure mode (overfit-on-rare-events, cross-sectional fee arithmetic, verification-loop instrument swaps, liquidation-data walls). Only 3 parameters (vol lookback, vol threshold percentile, trend filter length) keeps the optimization surface tiny against the 233-death overfit graveyard. The academic pedigree of vol-managed timing gives it genuine out-of-sample support rather than an in-sample-tuned curve.

IMPORTANT — the previous_code attached to this task is not this hypothesis's strategy. It is EthMomentumIgnitionContinuationLS, a 30-minute long-SHORT momentum-ignition book on the USD-M perp, with its own docstring about an iteration-1 COIN-M inverse-contract failure. This hypothesis is a daily, long-FLAT, volatility-managed spot timer. I could not make a minimal venue edit to that file and satisfy the reported failure: Layer 1.5 demands ETHUSDT.BINANCE_SPOT, and a long-short book on a CASH spot account has every SELL entry dropped by the base template's spot-short guard, so half its signals would be dead and its identity would still contradict the hypothesis. So I implemented the hypothesis as written and flag the mismatch rather than shipping a strategy that cannot be what it claims. The Layer-1.5 failure itself is fixed directly: instrument_id and bar_type are now ETHUSDT.BINANCE_SPOT with 1-DAY bars, leverage 1.0. Design follows the hypothesis exactly — Moreira-Muir volatility management, with realized vol ranked against its OWN trailing history (scale-free across eras rather than an absolute vol constant, which would not survive 2018 vs 2026 regimes), an SMA used only as a regime filter, and inverse-vol position sizing. It sits OUT during capitulation rather than into it, reads no supplementary data, and keeps exactly three searched parameters against the overfit graveyard. I validated the gate on the real series — 3,305 ETH spot daily bars from 2017-08 to 2026-09 — and at defaults it is invested 37.1% of days with 172 regime flips (~86 round trips, average hold ~14 days). That rules out both a zero-trade Layer-3 failure and the churn that would breach spot's 0.2% round-trip cost, since multi-week holds amortize it. Sizing only occurs while flat, so on a CASH account get_account_equity() is the full portfolio at that instant; quantities are floored to the instrument's own size precision with a min-notional check. Buffers are bounded and per-bar work is O(1) amortized, so there is no timeout exposure.

L49/L19 long-only vol-targeted crypto-beta family: beta-carried (beta 0.23, corr 0.56) with NEGATIVE information ratio (-0.449) versus its own buy-hold benchmark, so the vol-managed regime timing underperforms simply holding ETH on a risk-adjusted basis. Max drawdown 52.95% (CI to 75%) exceeds the 50% hard-abandon bar and falsifies the tail-truncation thesis (stressed tercile -52.5%). Sharpe 0.72 with CI low 0.063 straddles zero. Same signature as BtcSpotVolatilityTargetedLongAccumulation and VolTargetedLongCashBasket10Majors abandoned this session; no tuning of the 3 parameters converts a beta stream with negative IR and a 53% drawdown into a promotable edge — not worth 2 hours of optimization.

Long-flat volatility-managed regime timer on ETHUSDT.BINANCE_SPOT, daily bars, pure OHLCV. Each close it measures realized volatility as the stdev of the last vol_lookback daily log returns, ranks that value against its own trailing vol_rank_window history, and compares price to its trend_len-day SMA. It holds ETH only while the vol percentile is at or below vol_pct AND price is above the SMA — a calm uptrend regime — and otherwise sits in cash, so it exits into volatility spikes rather than buying them. Exposure is volatility-targeted: the invested fraction is target_vol / realized_vol clamped to [min_frac, max_frac] and hard-capped at 1.0x equity, so calm regimes get full size and choppier ones get scaled down. CASH account, long-only, no shorts, no leverage, no liquidation risk. Three searched parameters: vol_lookback, vol_pct, trend_len.

Verification failed (Layer 1.5 — hypothesis/config consistency) [class=hypothesis_mismatch]: - config.instrument_id is ETHUSDT.BINANCE but the hypothesis declares ETHUSDT.BINANCE_SPOT — same asset, DIFFERENT venue/contract type (fees, settlement and the mechanism's premise differ). Trade the declared instrument.

No fix required; verify full-history backtest hits the developer's measured ~86 round trips.

Risk-on requires ~trend_len (100) + min_obs (60) warm-up bars before the first entry can fire (legitimate, no look-ahead). Reduces trade count only on truncated windows; full history unaffected.

Clean, low-parameter (3 searched) single-instrument implementation on the correct BINANCE_SPOT venue, long-flat with no leverage — structurally sound and not a code defect

Positive alpha point estimate (0.15) and PSR 0.946 keep it from being obviously dead

max_drawdown 52.95% (CI high 75.1%) — above the 50% hard-abandon bar; the vol-off switch failed to truncate the tail it exists to cut (stressed regime -52.5%, ftx -17.5%, rate_shock -30.7%)

information_ratio -0.449 vs its own buy-hold benchmark (benchmark_meaningful=true): the regime timing SUBTRACTS risk-adjusted value versus simply holding ETH — the +706% headline is long-crypto beta (beta 0.23, corr 0.56), not timing skill

Sharpe 0.72 with bootstrap CI [0.063, 1.34] — low bound essentially at zero; edge indistinguishable from beta capture in the bull years

avg_trade_return_pct 5.45% is a compounding/position-riding artifact (avg_position_pct 84.9%), not real per-trade edge

Loses in the stressed vol tercile (-52.5%) — soft gate failure; the thesis (higher Sharpe / smaller tail in vol-managed exposure) is falsified on this instrument

EthSpotVolatilityManagedRiskOnOffLongFlatDaily

Verification failed (Layer 4 — QA review): - The strategy is structurally unmeasurable at the backtest horizon. The sandbox confirms it: over data_days=158 it produced total_trades=0 completed round-trips (trades_count=1 is a single position opened and never closed), metrics_reliable=false, and every performance metric returned 0.0. The config rationale itself estimates only ~13 round trips over ~400 daily bars. A daily long-flat regime timer that holds ETH continuously through low-vol uptrends is inherently very low frequency: the arm gate (vol percentile <= 0.60 AND close > SMA50) plus ~81-bar warmup (vol_pct_window=60 × 21-close vol samples) leaves few usable bars, and each hold spans many. ~13 trades is far below the ~100-trade floor needed to distinguish edge from noise; the analyst will abandon it as unmeasurable downstream — the exact low-fire-rate pattern that burns iterations. - Pure-OHLCV single-instrument strategies have a near-zero historical survival rate in this factory. The academic pedigree and non-overfit design are favorable but don't change the base rate. This is context for the analyst, not a code defect — the code faithfully implements the stated hypothesis. - Spot round-trip ~0.20% with a 0.25% avg_trade_return floor. Multi-day holds should clear it, but with so few trades a couple of boundary whipsaws near the SMA/vol threshold could dominate the average.

SpotMajorsVolatilityManagedRiskOnOffLongFlatDaily

Verification failed (Layer 4 — QA review): - The hypothesis is emphatically and repeatedly a SINGLE-INSTRUMENT strategy: the title says 'Single-Instrument ETHUSDT.BINANCE_SPOT' and the description frames the edge as vol-managed regime timing on ETH ONLY ('NOT a rank-rotation basket'). The code (iteration 2) instead runs the identical sleeve across an 8-name basket: config.extra_instruments = [BTC, BNB, SOL, XRP, ADA, LTC, DOGE].BINANCE_SPOT and on_extra_bar drives 7 additional sleeves. Per the QA checklist, every asset the strategy trades must be named in the hypothesis; 7 of the 8 traded instruments are NOT in the hypothesis. This converts a single-name timer into a diversified basket with a materially different risk profile. The hypothesis itself lists 'verification-loop instrument swaps' as a failure mode to avoid, and this iteration is exactly such an in-loop instrument expansion. The code no longer implements the hypothesis as written. - Informational (analyst domain, not a code defect): the sandbox already shows a decisively negative edge on a measurable sample — avg_trade_return_pct -2.13% (below the spot fee floor), profit_factor 0.32, win_rate 13.8%, total_return -9.5%, Sharpe -1.09. Entry (low-vol uptrend) and exit (trend break or vol spike) polarity correctly matches Moreira-Muir, so this is a timing/edge problem, not an obvious polarity bug. Likely to be abandoned downstream as edge-below-fees. - 65 completed trades over 363 days (~8 per sleeve) is below the ~100-trade measurability floor the developer targeted, so even the basket form is only borderline measurable over this window.

EthMomentumIgnitionContinuationLS