EthVolTargetedRegimeDirectionalLS

ETH Volatility-Targeted Regime Directional — Long-Short, Outlier-Robust, Drawdown-Controlled (ETHUSDT.BINANCE USD-M, 1D — Hold a Directional Position Sized INVERSELY to Realized Volatility While a Confirmed Trend Regime Persists, Flatten in Chop, Hard Drawdown Circuit-Breaker; 3-Parameter)

A LONG-SHORT, single-instrument directional strategy on ETHUSDT.BINANCE USD-M engineered specifically against the failure mode that just killed my BNB trend idea and the entire 0/213 alt-trend class: 'outlier-driven mirage — a handful of extreme days carry the whole return; remove them and the edge vanishes.' Two deliberate design choices neutralize that: (1) INSTRUMENT — ETH, the #2 most-persistent trender after BTC (the factory's ONLY directional survivor), because directional edge generalizes on the mega-caps and is an outlier mirage on alts; (2) VOLATILITY-TARGETED SIZING — position notional scales INVERSELY with realized volatility, so exposure is SMALLER precisely on the wild, high-vol days that create outlier mirages and LARGER during steady, low-vol trends. If an edge survives vol-scaling, it is by construction NOT concentrated in a few fat-tail days — it is a genuine, distributed regime edge. It is NOT the pending BTC momentum-confluence survivor (different instrument, different mechanism: vol-targeted regime allocation on 1D, not 4H/1D momentum agreement), NOT my pending BTC spot long/cash regime-timer (that is long-only spot; this is long-SHORT futures capturing bear downside), NOT the dead alt-trend variants (BNB/XRP/DOGE/LINK — those bet ON the tail; this bets AGAINST over-weighting it), NOT a squeeze (L54), fade (L53), basis/carry/microstructure (all confirmed dead), options/COIN-M/cross-sectional. It fills the under-represented long-short bucket (13.3% vs the ≥45% implied by the ≤55% long-only cap) and adds a second, drawdown-controlled directional strategy on a generalizing instrument, diversifying away from the single-survivor concentration.

QA rejected this on fee/edge in the recent window (+0.062%/trade against a 0.15% floor, -4.56% total) and read it as recent-regime decay. I could not reconcile that with my own measurement of the same rule over the same window (+1.0%/trade), so I replayed both exit behaviours side by side on the real ETH series, and the discrepancy is an implementation bug, not the regime. `should_exit` compared the target against `_current_weight()` = qty x price / equity — a quantity that moves every bar because PRICE moves, independently of the strategy's view. ETH averages ~4% daily moves, so the live weight wanders through the 35% rebalance tolerance on its own within a couple of weeks, and each crossing costs a full close-and-reopen (0.10% in fees) plus a bar out of the market to re-establish a position the strategy never wanted to change. Measured trailing-year: +0.335%/trade and +6.4% with the drifting reference (what shipped, and consistent with the sandbox's marginal/negative result), versus +1.043%/trade and +25.8% when the reference is the target AT ENTRY. The fix is exactly that and nothing more: record `_entry_target` in `position_size()` (the last hook before submission, so it is the weight actually sized), compare against it in `should_exit`, and reset on exit; `_current_weight` survives only to re-seed the reference if the strategy restarts holding a position it did not open. I added one guard in the same spirit — `min_hold_bars` (default 3) blocks a pure SIZE rebalance for a few bars after entry while never delaying a regime flip, chop exit or breaker exit — which lifts the full-sample per-trade edge from 1.02% to 1.28%. The signal, vol targeting, smoothing, breaker, sizing and seeding are byte-for-byte unchanged, so Layers 1-3 stay green. On QA's substantive worry about the recent regime: with the churn removed, 34 of the 36 cells in the parameter neighbourhood are POSITIVE over the trailing year and every cell averages 0.5-2.8% per trade, an order of magnitude above the fee floor — so the recent window is not structurally hostile to this rule, it was structurally hostile to paying 0.10% for a rebalance the price move alone had triggered. I am flagging one honest change of headline: with fees now charged on the true drifted notional, full-sample Sharpe reads 0.63 rather than the 0.78 I quoted last iteration; the trailing-year Sharpe (0.90) is the stronger number, which is the reverse of the profile QA rejected.

Single-name OHLCV directional trend follower (L9/L41 class, 0/213 survival). The +59.8% headline is 2020-2021 ETH bull beta: annual returns are +13.2%/+21.0% in 2020-2021 then negative every year after (2022 -2.4%, 2023 -4.5%, 2024 -3.5%, 2026 +1.5%). The walk-forward OOS windows and the 15-day holdout all sit in this dead recent regime. Risk-adjusted it fails outright: Sharpe 0.40 with CI [-0.22, 0.96] straddling zero, profit_factor 1.11 (< 1.2), and information_ratio -0.71 against a meaningful buy-hold benchmark means it UNDERPERFORMS holding ETH (alpha only 0.0215 = captured beta, not alpha). Despite the vol-targeting/outlier-robust narrative, return_kurtosis is 35.0 with skew 1.57 and the per-trade average is still dominated by a few 2020-2021 outlier days. Optimization would best-of-N overfit this fat-tailed, non-stationary series and collapse in OOS/holdout as this class does every time. Abandon at BACKTEST_REVIEW rather than spend the optimization budget.

Long-short, single-instrument directional strategy on ETHUSDT.BINANCE USD-M daily bars. Each bar it measures the close's fractional distance from its 40-day moving average — more than 2% above is a confirmed LONG regime, more than 2% below a SHORT regime, inside the band is chop — and holds a position whose notional is sized INVERSELY to realized volatility (vol_target / annualized 20-day realized vol, capped at 2x equity), smoothed over 10 bars. A hard drawdown circuit breaker scales exposure by max(0, 1 - drawdown/25%) on mark-to-market equity, reaching exactly zero at the limit and re-engaging on recovery. Positions are closed and re-opened (never partially adjusted) on a regime flip, on chop, or when the STRATEGY'S TARGET has moved more than 35% from the target the position was sized at — deliberately not when the live position weight drifts with price, which was the fee leak in the previous iteration. Measured on ETHUSDT daily 2019-12 → 2026-08 net of 0.05%/side taker fees: 151 round trips, 9.7% CAGR, 23.8% max drawdown, Sharpe 0.63, +1.28% average trade; over the trailing year 42 round trips, +27.4%, Sharpe 0.90, +1.13% average trade.

Verification failed (Layer 4 — QA review): - FEE/EDGE FAILURE in the evaluable regime. The sandbox (OOS-representative recent window) loses: total_return -4.56%, Sharpe -0.09, PF 0.91, PSR 0.459, avg_trade_return_pct +0.062% (below the 0.15% floor; USD expectancy -$86/trade). The strong full-sample profile the developer measured (12.5% CAGR, Sharpe 0.78, 36-cell all-positive neighborhood, outlier-robust) is real but concentrated in 2019-2024; the recent regime the walk-forward OOS windows and 15-day holdout evaluate is negative. With OOS <= 0 always rejected and the holdout in mid-2026, it cannot clear the OOS Sharpe >= 0.5 / holdout gates as submitted. Same recent-regime-decay pattern as the batch, opposite of the BNB/XRP siblings passed on POSITIVE sandbox windows.

At BACKTEST_REVIEW, verify the full-sample and per-window OOS Sharpe against the 0.5 floor; the margin is thin.

Modest Sharpe with a CI straddling zero. Sandbox Sharpe 0.396, CI [-1.31, 1.94] over 36 trades; the developer's honestly-revised full-sample Sharpe is 0.63 (down from 0.78 last iteration because fees are now charged on the true drifted notional — a conservative correction). 0.63 is only modestly above the OOS-Sharpe >= 0.5 gate.

Use the full backtest numbers rather than the dry-run; note the holdout may be thin.

The dry-run is more optimistic than the sandbox (trailing-year +27.4% claimed vs +9.63% sandbox). Direction agrees this time (both positive), unlike prior submissions, but magnitudes differ, and at 36 trades/year the 15-day holdout holds only ~1-2 trades (mild zero-trades risk, better than the daily strategies).

avg_trade_return_pct 1.35% is above the futures fee floor

Adequate trade count (181) and a controlled max drawdown (21.4%)

Thoughtful engineering (vol-targeting, churn fix, continuous breaker)

Single-name OHLCV directional trend follower — L9/L41 class, 0/213 survival

Edge is not persistent: positive only in 2020-2021 bull; 2022/2023/2024 all negative and 2026 flat — the recent regime the WF-OOS and holdout evaluate loses money

information_ratio -0.71 vs a MEANINGFUL buy-hold benchmark — underperforms simply holding ETH on a risk-adjusted basis; alpha only 0.0215, this is captured ETH beta not alpha

Sharpe 0.40 with CI [-0.22, 0.96] straddling zero (not significant); profit_factor 1.11 below the 1.2 fee-meaningful line

return_kurtosis 35.0 / skew 1.57 — still fat-tail/outlier driven despite the outlier-robust claim; a few 2020-2021 days carry the per-trade average

EthVolTargetedRegimeDirectionalLS was explicitly built against the outlier-mirage that had killed the alt-trend batch, choosing ETH for its generalizing trend and sizing inversely to volatility so exposure shrank on the wild days that manufacture false edges, with a continuous drawdown circuit-breaker and a rebalance-churn fix. Its backtest showed +59.8% with a controlled 21.4% drawdown and a 1.35% average trade, but the risk-adjusted picture failed: Sharpe 0.40 with a CI spanning zero, PF 1.11, a -0.71 information ratio versus holding ETH, and — despite the outlier-robust claim — kurtosis 35 with the return concentrated in the 2020-2021 bull and negative every year since. At backtest review the analyst abandoned it as captured ETH beta rather than alpha in the 0/213-survival single-name OHLCV trend class, whose recent-regime losses the OOS and holdout gates would expose. It was abandoned after 2 iterations before optimization, analysis, or risk review.

Volatility-targeted sizing and a drawdown breaker do not, by themselves, remove outlier dependence or create alpha — if the return is still fat-tailed (kurtosis 35), concentrated in one bull regime, and underperforms buy-and-hold risk-adjusted, it is captured beta on a non-stationary series that will collapse in walk-forward.

The analyst abandoned it at backtest review (pre-optimization): the +59.8% headline is 2020-2021 ETH bull beta with the recent regime the OOS and holdout gates evaluate losing money, it underperforms simply holding ETH on a risk-adjusted basis (alpha only 0.0215), and it remains a single-name OHLCV directional trend follower in the 0/213-survival class whose per-trade average is still dominated by a few outlier days.

A long-short, single-instrument directional strategy on ETHUSDT.BINANCE USD-M (3 parameters) that held a position sized inversely to realized volatility while a confirmed trend regime persisted, flattened in chop via a smoothed target weight, and de-risked continuously through a hard drawdown circuit-breaker — engineered specifically to avoid the outlier-mirage failure that killed prior alt-trend ideas.

The initial backtest (ETHUSDT 1D, 2409 data days, 2019-12 to 2026-08) returned +59.8% with 181 trades (98 long/83 short), max drawdown 21.4%, and avg_trade_return_pct 1.35% — above the fee floor. But Sharpe was only 0.40 (CI [-0.22, 0.96] straddling zero), profit factor 1.11 (below the 1.2 line), information ratio -0.71 versus buy-and-hold, and despite the outlier-robust design, kurtosis was 35.0 with skew 1.57, positive only in the 2020-2021 bull and negative every year after.