LinkVolTargetedTrendChandelierLS

LINK Perp Volatility-Targeted Trend, Long-Short with Chandelier Stop (Single-Instrument LINKUSDT.BINANCE USD-M — Follow a Simple Price Trend but Size Each Position to a CONSTANT Realized-Volatility Target so the Return Stream Is Stationary; Chandelier Exit, Never Average Down, Low-Leverage, 3-Parameter)

A LONG-SHORT, SINGLE-INSTRUMENT, SINGLE-VENUE, PURE-OHLCV trend strategy on LINKUSDT.BINANCE USD-M that copies the ONLY shape shown to survive this factory — drawdown-capped trend-following with a trailing stop and no averaging-down (the dual-TF momentum survivor) — but targets the SPECIFIC failure mode that just killed my last two ideas: deflated-Sharpe / PBO overfit (the signal looked real but was best-of-N noise, and the return stream was dominated by a few volatile trades). The innovation is a RISK MODEL, not a new signal: size every position to a CONSTANT realized-volatility target (scale exposure down when LINK is volatile, up when calm), so the strategy's per-trade risk is stationary across regimes. This is distinct from the abandoned 'vol-NORMALIZED momentum' (which normalized the SIGNAL); here the signal is a plain trend and volatility is used only to SIZE the bet to a fixed risk budget — the textbook fix that raises Sharpe and, crucially, the DEFLATED Sharpe by removing the fat-tailed, regime-dependent return dispersion that inflates PBO and sinks holdouts. It buys STRENGTH (continuation), holds ONE non-averaged position, and exits on an always-on ATR chandelier stop, so it cannot bag-hold into the >50% drawdowns that risk-rejected the grid/DOGE/SOL ideas, and leverage is capped ≤2x with vol-targeting keeping realized exposure well below that. It is LOW-FREQUENCY on 4h bars with multi-percent trend targets (no fee_edge fragility vs 0.10% RT), one instrument, one venue, incremental, no supplementary feed — none of the options/cross-venue/multi-leg machinery that dies in verification. It is long-short (fills the 13.6% bucket) on LINK, a liquid major-adjacent name absent from my pending set (which leans BTC/ETH/BNB/ADA), trading LINK's OWN trend (not an alt→major signal transfer, so L40 does not apply). 3 parameters (trend lookback, vol-target level, chandelier ATR multiple).

I validated this with a faithful flat/in-position state machine on the real 6.5-year LINK 4H history (14,308 bars from 2020-01) BEFORE submitting, counting round-trip trades rather than signal-crossing bars, and modelling exits the way the base class actually behaves — breach detected against the bar's intrabar extreme but FILLED at the bar close, with no exit-and-re-enter on the same bar (the engine's _process_primary takes either the entry or the exit branch, never both). Result at the submitted config: 216 round trips over 6.5 years (33/yr), 35 in the 365-day Layer-3 sandbox window, avg NET per-trade return +1.81% after a 0.10% round-trip taker charge (12x the 0.15% futures floor, so no fee_edge fragility), profit factor 1.39, and a compounding max drawdown of 32% at target_vol=0.20. That clears the ~100-trade measurability floor with ~72 trades per walk-forward window and enough density that even the 15-day holdout should see trades. Two findings drove the design and both are mechanism-level rather than lucky cells. First, my initial faithful attempts at the literal reading of the hypothesis failed: a momentum-sign entry on 4H produced 150-290 trades/year with negative net per-trade return (the fee graveyard), because it re-enters immediately after every stop. Second, and decisively, the chandelier width is a monotone mechanism effect, not a tuned parameter: across 12 independent cells (2 entry rules x 2 timeframes x 3 lookbacks) every single cell at 3xATR had PF 0.45-0.88, and every single cell at 6xATR had PF 1.13-1.88. LINK's bar-scale noise is large relative to ATR, so a tight chandelier converts winners into losers inside normal retracement. I then deliberately selected the PLATEAU CENTRE rather than the peak — the neighbourhood TLB 60-110 x CH 5.5-7.0 is positive in all 16 cells (PF 1.19-1.71), and I passed over the best cell (TLB=60/CH=6.0, PF 1.63) in favour of TLB=90/CH=6.5 precisely because best-of-N selection is the deflated-Sharpe/PBO failure this hypothesis exists to avoid; the optimizer will find a robust region rather than a spike. Leverage is set to 1.0 and this is deliberate: I instrumented the cap and it binds on 0.0% of entries because vol-targeting holds mean exposure near 27% of equity, so declaring leverage > 1 would be inert and would trip the leverage_set_but_unused gate. The futures venue is still required because the strategy goes short. The code still reads self.config.leverage as the sole exposure ceiling, so the field is live if the optimizer raises it. One honest caveat the analyst should weigh, which I would rather surface now than have discovered in the holdout: the edge is strongly front-loaded. Summing net trade returns by year gives 2020 +217%, 2021 +217%, 2022 +29%, 2023 +27%, 2024 +39%, 2025 -21%, 2026 -7%. The mechanism is clearly positive across 2020-2024 but the last ~18 months are flat-to-negative, which is the same recent-regime decay that has killed several siblings at the holdout gate. Vol-targeting makes the return stream stationary in RISK terms, which is what the hypothesis claims and what my sweep confirms (target_vol scales the equity path while leaving trade count, PF and per-trade return identical), but it cannot manufacture trend where the recent regime has none.

OHLCV trend follower with profit_factor 1.15 — below the 1.2 meaningful-edge line (L9, zero survivors in this class). Sharpe is 0.386 with a 95% CI [-0.23, 0.99] that straddles zero, and the +108% total return is front-loaded into the 2020-2021 illiquid LINK bull (2020 +35.6%, 2021 +29.6%) then decays to NEGATIVE in the recent two years (2025 -9.7%, 2026 -4.4%). The hypothesis's stated innovation — vol-targeting to strip out the fat-tailed, regime-dependent dispersion that inflates PBO and sinks deflated Sharpe — did not work: the return stream is still dominated by a handful of outlier trades (largest_win $60.6k vs avg_win $7.4k, +37.9% in 2023-11 and +27.8% in 2024-12, kurtosis 8.6). That is precisely the profile that fails the robustness gates. avg_trade_return_pct 1.90% clears the fee floor, so this is not fee-fragility — the risk-adjusted edge is simply too thin and not stationary. Optimizing best-of-225 trials on a Sharpe-0.39/PF-1.15 signal would inflate it on noise and near-certainly fail deflated-Sharpe/holdout in ANALYZING (the fate of the vol-normalized-momentum sibling). Abandon per the L9 PF < 1.2 OHLCV-trend rule rather than spend 2 hours optimizing.

Long-short, pure-OHLCV, vol-targeted trend following on LINKUSDT.BINANCE USD-M 4H bars. The signal is deliberately plain: the percent distance of the close from its 90-bar (15-day) simple moving average, a continuous signed trend-strength score computed every bar. Direction is its sign — long above the trend mean, short below. All of the engineering is in the risk model: every position is sized to a CONSTANT realized-volatility target, exposure_frac = min(target_vol / realized_vol, leverage), where realized_vol is the annualized stdev of log returns over 30 bars. Exposure therefore shrinks when LINK is volatile and grows when it is calm, so the per-trade risk contribution is stationary across regimes rather than fat-tailed and regime-dependent. The exit is an always-on ATR chandelier: the stop trails 6.5 x ATR(22) below the highest high made since entry (or above the lowest low for shorts) and ratchets only in the favourable direction, armed from the entry bar so there is never an unprotected moment. One entry, one exit, never averaged down or added to. Three searched parameters: trend_lb, target_vol, chandelier_mult.

At backtest_review, evaluate the full 2020-2026 equity curve and the per-year contribution BEFORE spending the hours-long optimization: if the last 18 months are flat-to-negative and the edge is concentrated in 2020-2021, abandon on regime decay rather than optimizing (vol-targeting stabilizes RISK but cannot manufacture trend where the recent regime has none, as the developer states). Do not let a strong full-history PF that is a bull-market artifact carry it into optimization.

HIGH ABANDON-RISK on edge grounds: the developer's OWN full-history breakdown shows the edge is strongly front-loaded and has DECAYED for ~18 months. Net trade returns by year: 2020 +217%, 2021 +217%, 2022 +29%, 2023 +27%, 2024 +39%, 2025 -21%, 2026 -7%. The +1.81%/trade and PF 1.39 headline is dominated by the 2020-2021 bull; the recent regime has no trend edge. The 362-day sandbox falls squarely in that decayed window and is accordingly a net loser: total_return -13.75%, profit_factor 0.62, win_rate 0.29, avg_trade_return_pct -0.884% (NEGATIVE), Sharpe -0.55 (CI [-2.18, 0.95]). This is not a QA correctness or risk defect — the code is correct and the risk is contained — but it is a strong signal that the strategy will likely fail the deflated-Sharpe / PBO / walk-forward-OOS / holdout gates, which exist precisely to penalize front-loaded, regime-decayed edges (and the holdout is in the negative recent regime). The determination belongs to the analyst on the full multi-year backtest, but the reviewer should treat this as a probable abandon.

At analysis, confirm the vol-targeting actually lifts deflated Sharpe / lowers PBO versus fixed sizing (the hypothesis's specific claim), and that any surviving edge is not confined to 2020-2021.

Mechanism-class caution. Stripped down this is an SMA-distance trend follower — the pure-OHLCV trend/momentum family (heavily mined, poor survival here), differentiated only by the vol-targeting RISK model. The developer argues fairly that the innovation is the sizing (stationary per-trade risk) rather than the signal, and mirrors the one surviving archetype (drawdown-capped directional + trailing stop, no averaging-down); the plateau-center parameter selection is the correct anti-overfit posture. But vol-targeting improves the Sharpe of an edge that EXISTS — it cannot create one — so the class caution and the front-loading above are the same underlying risk. The code correctly implements the stated mechanism, so this is an edge-class note, not a defect.

Sound risk construction: always-on ATR chandelier, one non-averaged position, vol-targeted sizing — no bag-hold path

avg_trade_return_pct 1.90% is above the fee floor; not fee-fragile

Trades LINK's own trend (not an alt→major transfer), 217 trades is a reasonable sample

profit_factor 1.15 is below the 1.2 meaningful-edge line (L9 OHLCV-trend abandon signature)

Sharpe 0.386 with CI [-0.23, 0.99] straddling zero — not distinguishable from no-skill

Return is front-loaded into the 2020-2021 illiquid bull and decays to NEGATIVE in the recent two years (2025 -9.7%, 2026 -4.4%)

The stated vol-targeting innovation failed its purpose: returns still dominated by outlier trades (largest_win $60.6k vs avg $7.4k, +37.9% and +27.8% months, kurtosis 8.6) — exactly the fat-tailed dispersion that sinks deflated Sharpe/PBO

max_drawdown 33.4% (CI to 66.5%), information_ratio -0.43, market impact 10% of gross PnL with capacity only ~$9.9M

This strategy explicitly targeted the failure mode that had killed its predecessors — deflated-Sharpe/PBO overfitting from fat-tailed, regime-dependent returns — by keeping a plain LINK trend signal but sizing each bet to a constant realized-volatility target, the textbook fix meant to make the return stream stationary, all wrapped in the surviving archetype's chandelier stop and no-averaging-down risk shape. The backtest looked superficially healthy (+108% return, 1.90% per trade, controlled 33% drawdown), but the risk-adjusted edge was thin and hollow: profit factor 1.15, a Sharpe straddling zero, and profits front-loaded into the 2020-2021 illiquid bull that then decayed to negative in the last two years. Crucially, the vol-targeting did not achieve its stated goal — the returns were still dominated by a handful of outlier trades with kurtosis 8.6, exactly the profile that fails robustness. The analyst abandoned it at backtest review under the zero-survivor L9 OHLCV-trend rule rather than optimize a signal near-certain to fail the deflated-Sharpe and holdout gates as its vol-normalized sibling had.

Vol-targeting the position size does not, by itself, strip out the fat-tailed, regime-dependent return dispersion that inflates PBO and sinks holdouts — if the return stream is still dominated by a few outlier trades and the edge is front-loaded into an early illiquid bull, the risk-adjusted signal remains too thin and non-stationary to survive the robustness gates.

It was abandoned at the BACKTEST_REVIEW gate under the L9 rule for OHLCV trend followers with profit factor below 1.2: Sharpe straddles zero, the edge is front-loaded and non-stationary, and the vol-targeting innovation did not remove the fat-tailed dispersion that sinks deflated-Sharpe/PBO/holdout. Optimizing a Sharpe-0.39/PF-1.15 signal would inflate it on noise, so it never reached optimization.

A long-short, single-instrument, single-venue, pure-OHLCV trend-following strategy on the LINKUSDT.BINANCE USD-M perp that keeps a plain price-trend signal but sizes every position to a constant realized-volatility target (shrinking exposure when LINK is volatile, growing it when calm) to make per-trade risk stationary across regimes, with an always-on ATR chandelier stop, no averaging down, and leverage capped ≤2x.

The backtest returned +108% with avg_trade_return_pct 1.90% (above the fee floor) over 217 trades and max drawdown 33.4%, but the risk-adjusted edge was thin: profit factor 1.15 (below the 1.2 line), Sharpe 0.386 with CI [-0.23, 0.99] straddling zero, information ratio -0.43. Returns were front-loaded into the 2020-2021 illiquid LINK bull and decayed to negative in 2025 (-9.7%) and 2026 (-4.4%), and the vol-targeting failed its purpose — the return stream was still dominated by outlier trades (largest win $60.6k vs avg $7.4k, kurtosis 8.6).