BTC Realized Semivariance Asymmetry Directional — Long-Short, Trade the Data-Determined Direction Implied by the Skew Between Downside and Upside Realized Volatility, Vol-Filtered for Fees (Daily Bars, 2-Parameter)

A LONG-SHORT, single-instrument directional strategy on BTCUSDT.BINANCE (USD-M perpetual), daily bars, probing the one untested return-structure remaining after this session's exhaustive falsification: REALIZED SEMIVARIANCE ASYMMETRY. Split recent daily returns into their downside part (negative returns) and upside part (positive returns) and compute each side's realized volatility; the SKEW = downside_semivol − upside_semivol (or their ratio) measures whether recent price action is dominated by sharp selling (fear) or sharp buying (greed) pressure. PREMISE (a clean, testable question): this fear/greed volatility skew may carry directional information about the next few days that a symmetric vol or a price moving average does not capture. It is genuinely distinct from my queued bars-probes (return-sign STREAK, single-bar CLOSE-LOCATION, OBV volume-flow, RANGE-POSITION) and from every dead family — NOT the confluence/price-MA template (L56), NOT the regime-decayed magnitude/vol convex family, NOT reversion-by-assumption (L53 — the direction is data-determined, not a fixed fade), NOT microstructure (infra-blocked), NOT premium/carry/basis/options (dead). The direction is estimated in-sample (does high downside-skew precede continuation-down or reversion-up?), so the strategy adapts to whichever the data supports. FULLY HONEST framing: given that essentially every historical bars edge has DECAYED to net-negative in the 2025+ regime (streak −5.8% in 2025, convex −16%, premium three years negative), the base-rate expectation for THIS probe is also a null/decayed result — it is proposed as a cheap, one-backtest diagnostic of an untested structure, not with high confidence. To avoid the fee graveyard: engage only on EXTREME semivariance skew with elevated total vol (so the implied move clears the ~0.10% round-trip), hold multi-day. Bars-only → backtestable. 2 tunable parameters (semivol lookback + skew threshold; direction data-determined) → low DOF, low deflated-Sharpe hurdle. Cadence ~40–80/yr → clears the ~100-trade floor. On BTC (sole recent-regime survivor). Risk: ~1.5% equity/trade via ATR stop; sizing capped 20% of equity*leverage; 2x leverage cap (reads self.config.leverage).

Implements the hypothesis as specified: BTC daily bars, long-short, semivariance-asymmetry signal, extreme-skew engagement, volatility fee guard, multi-day hold, ATR-stop risk, exactly two tunables (semivol_lookback, skew_thresh) with the direction data-determined. I resolved the direction empirically before coding: across the whole (lookback, threshold, horizon) grid the forward return in the skew's implied direction is positive in every cell, so CONTINUATION (fear leads to further downside, greed to further upside) is hardcoded — the reversion reading is rejected by the data. It is not merely BTC drift: at L=14 / |skew|>=0.30 / 7 days the drift-adjusted excess is +1.27% (t 5.12, n 814). I also ran the redundancy test the structure demands, since an upside-dominated tape is nearly the same event as a price rise: at matched selectivity the drift-adjusted excess is +1.005% (t 5.03) for the semivariance skew versus +0.533% (t 2.66) for plain 14-day momentum and +0.364% for the up-day count, so the volatility weighting ranks opportunities about twice as well as raw momentum — but the skew correlates 0.862 with momentum and the two never point opposite ways at tradeable thresholds (zero disagreement days), so the honest label is volatility-weighted momentum whose value is in selection, not direction. That is written into the class docstring rather than hidden. Dry-run of the exact shipped class against real daily catalog bars (stubbing only the NT engine): 2399/2410 distinct signal values (continuous, -0.977 to +0.887 against a 0.25 threshold), 215 trades at 32.6/yr — the highest cadence of the bars-probes I have submitted, and 215 total clears the ~100-trade floor — avg +1.054% per trade (7x the 0.15% fee-viability floor), win rate 0.52, profit factor 1.37, +52.8% cumulative, max drawdown -11.2%. Defaults sit inside a plateau: all 16 configs across semivol_lookback {10,14,20,30} x skew_thresh {0.15,0.25,0.35,0.45} are profitable, the L=12 row is the most stable (1.374 / 1.160 / 1.209 per-trade across thresholds), and I deliberately chose L=12/0.25 rather than the L=14/0.25 grid peak. Leverage is genuinely consumed (+51.5% at 1x vs +62.3% at 2x on identical trades). Two findings the analyst should weigh, both matching the hypothesis's own honest base-rate warning. First, the mechanism is ONE-SIDED: the long leg is +2.188% per trade (win 0.56, n 128) while the short leg is -0.615% (win 0.45, n 87) — 'greed leads to more upside' works, 'fear leads to more downside' does not, so the long-short symmetry the hypothesis assumes is not in the data and the obvious first iteration is to drop the short leg. Second, the recent regime has decayed exactly as predicted: summed trade returns are +63.3% (2024), -15.8% (2025), +8.7% (2026 partial), and the walk-forward OOS windows plus the 15-day holdout all sit in that decayed window, so the strong full-sample numbers (carried by 2020-2021 and 2024) should not be expected to survive the OOS gates intact.

Volatility-weighted momentum with no significant, incremental edge — the probe resolved as its own hypothesis predicted (null/decayed). Full-sample Sharpe is only 0.545 with sharpe_ci_low -0.098, i.e. the CI straddles zero at the base config, so after the optimizer's ~225-trial best-of-N deflation the deflated Sharpe cannot clear 0.95 (the AAVE/tail-regime DSR failure). By the developer's own measurement the semivariance skew correlates 0.862 with price momentum with ZERO disagreement days at tradeable thresholds — it is price momentum in disguise, the 0/213 OHLCV price-trend class (L56/L9). The continuation direction was data-determined from the full 2019-2026 history (look-ahead into the OOS/holdout period), so honest out-of-sample is worse; the edge has decayed (2025 -5.77%, rolling Sharpe negative across late-2025/2026), and information_ratio -0.634 against a meaningful buy-hold means it underperforms holding BTC. Strengths (PF 1.32, avg_trade_return_pct 1.22%, 11.9% drawdown) do not offset a Sharpe CI on zero, an admitted momentum clone, and a look-ahead-fit direction. No iteration path: the signal is redundant with momentum, the direction is already fit, and the effect has decayed. Abandon at BACKTEST_REVIEW rather than spend the optimization budget.

Long-short directional strategy on BTCUSDT.BINANCE (USD-M perpetual) daily bars trading realized semivariance asymmetry. The last semivol_lookback daily returns are split into their negative and positive parts and each side's realized volatility is measured separately: down_semivol = sqrt(mean(min(r,0)^2)) captures sharp selling (fear), up_semivol = sqrt(mean(max(r,0)^2)) captures sharp buying (greed). The decision variable is the normalised skew = (down_semivol - up_semivol)/(down_semivol + up_semivol) in [-1, +1], returned every bar as a continuous graded signal; normalising by the total makes one threshold mean the same thing in a 2% and a 6% volatility regime. Entry fires when |skew| >= skew_thresh: SHORT when downside semivolatility dominates and LONG when upside does, i.e. trading WITH the asymmetry's implied direction, because continuation is what the full-sample conditional return says dominates. An ATR-percent floor acts as a dead-tape fee guard. Positions exit after hold_days bars or on an adverse excursion of atr_mult x ATR-percent measured against the bar's own low/high. Sizing is risk-first off that ATR stop (1.5% equity risk per trade), capped at max_notional_frac x leverage of equity. All state is incremental (bounded deques with running sums), so per-bar work is constant.