LinkWeekly12BarBreakoutAtrTrailLong

LINK Weekly 12-Bar High Breakout Long with ATR Trailing Stop

A LONG-ONLY weekly-bar momentum breakout on LINKUSDT.BINANCE perpetual futures (Binance USD-M). Entry on weekly close that exceeds the highest weekly close of the prior 12 weeks (quarterly breakout). Exit via ATR-based trailing stop OR a 26-week time exit. CRITICAL EMPIRICAL GROUNDING from this session: the XRP pullback strategy just failed catastrophically (-7.15 Sharpe, 19% win rate, contrarian signal), confirming that PULLBACK and MEAN-REVERSION mechanisms do NOT work on crypto perpetuals. The empirically working pattern across this session has been pure MOMENTUM/BREAKOUT (BTC ATR-compression range expansion, volume breakouts, trend continuation — all moved through the pipeline successfully). This hypothesis sticks rigorously to the WORKING pattern: zero pullback, zero reversal, zero contrarian elements. Pure 'price breaks above 12-week high = enter, ride the trend, exit on trailing stop.' LINK is selected because: (1) confirmed in the precision map (already running daily trend continuation in pipeline), (2) ZERO existing LINK strategies use weekly bars — adds within-instrument diversification, (3) LINK has 5+ years clean Binance USD-M data with no deprecation risk, (4) LINK's oracle-infrastructure positioning produces episodic multi-month trends (Q3 2021, Q4 2023, Q1 2024) that weekly breakouts cleanly capture. Structurally distinct from the 13 pipeline strategies: NOT pullback (avoiding the now-falsified class), NOT cross-asset, NOT supp-data, NOT calendar, NOT accumulation, NOT options. The fresh dimensions vs pipeline: WEEKLY timeframe (only DOGE Triple Screen partially uses weekly), 12-bar lookback (none of the 50-day, 4H, daily-20 lookbacks in pipeline match), ATR-trailing exit (different from fixed-stop and MA-crossover exits used elsewhere). Simple parameter set: lookback=12, ATR_period=14, trailing_multiplier=3.0 — only three free parameters, minimizing overfitting surface area.

Iteration 4 reconciles the two blocking issues by collapsing to the single LINK instrument the approved hypothesis specifies. (1) Exposure cap: with exactly one position, sizing anchors to get_account_equity() (the venue's stable USDT balance, NOT cash + unrealized mark-to-market) and caps single-position notional at equity * max_notional_frac * leverage, so realized exposure stays <= ~100% on the leverage=1 account — fixing the prior cash+gross ceiling that inflated to 455%. (2) Code/config now match the hypothesis exactly: LINKUSDT.BINANCE USD-M perp, weekly bars, 12-week breakout, ATR trailing + time exit, pure momentum with zero pullback/mean-reversion/contrarian elements. The basket, extra instruments, and cross-asset gross accounting are removed entirely.

Not worth optimizing: only 6 trades in ~3.5 years is too small a sample for the 225-trial sweep to produce anything but noise-fit parameters. The strategy loses money at base (PF 0.97, negative expectancy, 16.7% win rate) with the entire positive return carried by a single 2020-11 outlier (kurtosis 608, tail_ratio 0.05) while every subsequent year is negative, and it posts negative alpha/information-ratio vs its own LINK buy-hold. The code faithfully implements the hypothesis (clean long-only breakouts, no mechanism mismatch), so this is not a developer-fixable bug (not iterate) — the weekly 12-bar breakout premise simply generates too few, money-losing signals on LINK to support a robust edge, and no parameter region exists to tune toward.

Single-asset LONG-ONLY weekly quarterly-breakout momentum on LINKUSDT.BINANCE (USD-M perp). Enters when the weekly close breaks above the highest weekly close of the prior 12 weeks; exits on a 3x-ATR trailing stop off the highest high since entry, or a 26-week time stop. Position sized by 2% ATR risk, hard-capped so single-position notional stays <= 95% of stable equity (leverage=1).

Code faithfully implements the hypothesis: 6 long-only quarterly breakouts, zero shorts, no contrarian/pullback element

Simple 3-parameter surface, low overfitting surface area by design

Position sizing is sane (avg_position_pct 3.8%, max_drawdown 9.3%)

Only 6 trades in ~3.5 years — far too few for meaningful optimization; the optimizer would fit noise

Loses money: profit_factor 0.97, negative expectancy, win_rate 16.7% (1 of 6)

Entire positive return is a single 2020-11 outlier (return_kurtosis 608, tail_ratio 0.05); every year after 2020 is negative

Negative alpha (-0.013) and negative information_ratio (-0.74) vs its own LINK buy-hold

Near-zero Sharpe 0.005 with CI [-3.64, 0.82] straddling zero — no distinguishable edge

Do NOT optimize — two blocking problems. (1) The gross-exposure cap is still broken: realized exposure_pct = 455% on a leverage=1 BINANCE_SPOT CASH account, which is impossible to hold. Root cause: _sized_qty builds its ceiling from equity_ref = cash + gross (mark-to-market of open positions), so in a bull run the marked-up book inflates max_gross faster than cash is spent. On a CASH account the budget must be anchored to actual FREE CASH balance, not cash + unrealized MTM. Fix so reported exposure_pct stays <= ~100%. (2) Hypothesis says single-asset LINK weekly on LINKUSDT.BINANCE (USD-M perp), but the code is a 10-asset long-only basket on BINANCE_SPOT. Reconcile with the Research Lead and make code + config match the approved hypothesis. After both fixes, re-backtest and confirm metrics are plausible (Sharpe will drop sharply once leverage is real) before resubmitting. See workspace/discussions/7f60e2e2-6033-42cd-a83a-6527386e2174/iteration_4_feedback.md.

Grounded in the session's observation that pullback and mean-reversion had failed on crypto perps, this strategy stuck rigorously to pure momentum: a long-only LINK weekly breakout above the 12-week high with an ATR trailing stop. The iteration-4 code correctly collapsed a prior broken 10-asset spot basket down to the single LINK perp the hypothesis specified, and it ran faithfully — but produced only 6 trades in ~3.5 years, lost money (profit factor 0.97, 16.7% win rate), and owed its entire positive return to a lone November-2020 outlier while every year after was negative. The reviewer abandoned it before optimization, concluding the weekly 12-bar breakout premise simply generates too few, money-losing signals on LINK to support a robust edge, with no parameter region worth tuning toward.

A weekly 12-bar breakout on a single instrument produces too few signals (6 in 3.5 years) to establish a robust edge, and a positive headline return driven entirely by one outlier bar with every subsequent year negative is not a tradeable edge — signal frequency must be sufficient for the sample to be meaningful.

The backtest reviewer issued an 'abandon' verdict at the pre-optimization gate: only 6 trades in ~3.5 years is far too small a sample to optimize without fitting noise, and with a losing base result and no positive parameter region to tune toward it never advanced to optimization or analyst review.

A long-only weekly momentum breakout on LINKUSDT.BINANCE USD-M perps that entered when a weekly close broke above the highest close of the prior 12 weeks (quarterly breakout) and exited on a 3x-ATR trailing stop or a 26-week time stop, deliberately avoiding the pullback/mean-reversion mechanisms that had failed earlier in the session.

The code implemented the hypothesis cleanly — 6 long-only breakouts, zero shorts — but generated too few, money-losing signals: profit factor 0.97, negative expectancy (-$29.56/trade), 16.7% win rate (1 of 6), near-zero Sharpe 0.005 (CI [-3.64, 0.82]), and negative alpha (-0.013) and information ratio (-0.74) versus its own LINK buy-hold, with the entire positive return carried by a single Nov-2020 outlier (kurtosis 608, tail_ratio 0.05).

LinkBasketWeekly12BarBreakoutAtrTrailLong