BtcAtrChannelSuperTrendRegimeFilteredLS

BTC Volatility-Adaptive ATR-Channel Trend (SuperTrend-Style Stop-and-Reverse) — Long-Short, Single-Instrument OHLCV: Trend Direction Defined by an ATR-Band Flip, Filtered by a Slow-EMA Regime to Kill Chop, 4H Bars (BINANCE BTCUSDT USD-M, 3-Parameter)

A LONG-SHORT, SINGLE-INSTRUMENT, pure-OHLCV trend-following system on BTCUSDT.BINANCE USD-M perpetual built on a VOLATILITY-ADAPTIVE ATR CHANNEL (SuperTrend / Chandelier construction), which is mechanically DIFFERENT from the factory's one survivor (BTC dual-timeframe MOMENTUM confluence, which reads two momentum series and goes flat on disagreement). Here the trend direction is defined by the ATR band itself: price closing beyond an ATR-multiple band flips the position, and the band then trails the position as a volatility-scaled stop — the signal, the entry, and the trailing exit are one integrated object, not separate momentum indicators. To avoid the whipsaw that kills naive always-in stop-and-reverse systems in ranges, a single SLOW-EMA regime filter permits LONG flips only when price is above the slow EMA and SHORT flips only when below (flat/held otherwise) — this is the selectivity that made the winner robust, expressed through a volatility channel rather than a momentum-confluence, so it is neither the L62-banned dual-timeframe-momentum clone nor any of the pending ideas (ETH mean-reversion, ETH/BTC regime-agreement trend). It targets the exact gates that have been killing hypotheses: pure OHLCV with FULL 2020-2026 history and a naturally high flip count (~200-400 trades) that clears the >100-trade measurability tripwire (which killed the 2023+-feed BNB idea); a strong-major traded instrument for a high base-Sharpe prior (weak supplementary/positioning/funding overlays all became net losers or deflated to DSR≈0); and only 3 wide-plateau parameters (ATR period, ATR multiplier, slow-EMA length) — standard, robust, non-cliff values — to survive the ~225-trial deflated-Sharpe test. Single-instrument, no supplementary feed, no second leg, so it runs cleanly end-to-end in the Layer-3 sandbox (unlike the carry/cross-venue class that dies in verification per L69).

Iteration 2 fixes exactly the two things QA named, with the smallest possible change: no logic, no imports, no structure touched — only the primary timeframe (4H -> 2H), two parameter defaults, and the ema_period clamp/bounds. (1) TRADE COUNT / SIGNIFICANCE. QA: ~27/yr, 175 total, 21 in the sandbox, whole-sample t-stat 1.67 below the ~2.0 bar, sandbox Sharpe CI straddling zero. The same channel run on 2H bars at atr 14 / mult 3.5 / EMA 400 gives 328 trades (~51/yr) — inside the hypothesis's promised 200-400 band and ~2.5x the sandbox sample (about 52 trades instead of 21) — with a HIGHER whole-sample t-stat of 2.48, profit_factor 1.57 and +1.11% mean net per trade (over 10x the 0.10% round-trip fee, so the faster timeframe does not eat the edge; 1H does — I tested it and every 1H cell collapses to PF 1.03-1.09, which is why I stopped at 2H). (2) SPIKE CONCENTRATION. QA: 'remove the top 1-2 trends and it is break-even.' On this cell, deleting the two largest winners still leaves +261.8 of +355 points (74% of PnL), and the book is positive in six of seven calendar years (only 2025 negative; the last 12 months are +36.4 over 52 trades). (3) MONOTONE-TO-BOUNDARY. QA: the optimizer will run to the upper clamp. On 2H the surface now has an INTERIOR optimum on every axis: atr_mult 2.5/3.0/3.5/4.0/4.5/5.0 -> t 1.85/1.53/2.48/2.37/2.16/1.67 (peak at the submitted 3.5, DECLINING into the 5.0 clamp), ema_period 150/200/250/300/400/500 -> t 1.61/1.64/1.81/2.15/2.48/2.43 (broad plateau 300-500, submitted 400 sits inside it), atr_period 7/10/14/21 -> t 2.12/2.72/2.48/2.20 (nearly inert). The regime filter remains genuinely additive rather than decorative: removing it drops PF 1.57 -> 1.25 and t 2.48 -> 1.48. Remaining caveat for the Analyst, stated plainly: this is still a single-name pure-OHLCV trend construction, i.e. the class QA flagged on class priors, and no amount of re-parameterisation changes that — what iteration 2 can honestly claim is that the specific statistical objections (thin sample, sub-2 t-stat, outlier dependence, boundary-seeking parameters) are now all measurably answered.

backtest_review_stuck_3_timeouts

Long/short BTCUSDT.BINANCE USD-M perp on 2H bars. A SuperTrend-style ATR channel defines the trend: bands at (high+low)/2 +/- atr_mult * ATR(atr_period) ratchet one-way until price closes through them, flipping the channel direction; the active band is then the volatility-scaled trailing stop. Every bar the strategy returns the continuous distance from that band in ATR units (sign = channel direction, magnitude = stop distance). It goes LONG on an up-flip only while close > EMA(400) and SHORT on a down-flip only while close < EMA(400); when the channel flips into a direction the slow EMA does not endorse it stands flat. The exit is the channel itself — the position closes when the direction flips against it, with no separate stop layered on top. All state is incremental (Wilder ATR, one EMA accumulator, two band values), so per-bar work is O(1). Sizing risks 2% of equity across the actual band distance, capped at 60% of equity notional, leverage 1.0.

Verification failed (Layer 4 — QA review): - The hypothesis's OWN survival criterion is undelivered, and the submitted config will not clear the deflated-Sharpe gate (L7/L9/L16/L41). The hypothesis was explicitly pitched to survive the ~225-trial deflated-Sharpe test via a HIGH flip count ('200-400 trades', >100/yr) to tighten the Sharpe CI. The submitted config instead trades ~27/yr — 175 over the full history, only 21 in the sandbox — because the developer traded count for per-trade edge. The consequences: sandbox Sharpe 0.676 with sharpe_ci_low -1.07 (CI straddles zero) over 21 trades, and the developer's OWN full-history t-stat is 1.67 — below the ~2.0 significance bar BEFORE any deflation, so over ~225 optimization trials the Deflated-Sharpe will fall well under the floor. The result is carried by a few big trend wins (avg_win $2976 vs avg_loss $1065, win_rate 0.38, largest_win $7597 a large fraction of PnL) — remove the top 1-2 trends and it is break-even, the classic thin-sample trend-follower whose edge does not survive OOS. - This is a textbook pure-OHLCV single-name trend clone — the 0/213 OHLCV / 0/97 trend-clone dead class (L7/L9/L41). A SuperTrend/Chandelier ATR-channel with a slow-EMA regime filter is a standard trend-following breakout construction (adjacent to the abandoned Donchian/breakout class); recent siblings in this exact class were abandoned with the identical signature (SolOpenInterestConfirmedBreakout Donchian, HyperliquidEthTrendFundingCrowdingBrake EMA-trend — both base Sharpe CI straddling zero, spike-concentrated). Compounding the overfit risk, the developer concedes the parameter response is 'close to monotone' toward the upper clamp (mult 5.0/EMA 300 -> PF 2.41 on only 111 trades), so the optimizer will run to the boundary and land on an even-thinner, more concentrated cell — the opposite of the 'wide robust plateau' the hypothesis requires. A monotone-to-boundary surface is not a plateau.

Analyst: treat the DSR and OOS-Sharpe on the full 2020-2026 sample as decisive. Note the developer's honest caveats — 2025 was negative, and 1H bars collapse the edge (PF 1.03-1.09), so 2H is a real boundary of the edge, not a free knob.

Pure-OHLCV single-name trend construction (L7/L9/L41 0/213 prior): the deflated-Sharpe over the actual ~225 optimization trials is the binding, un-waivable test, and it is computed downstream, not here. The developer's supporting numbers are strong and — unlike iteration 1 — now plausibly deflation-surviving: full-history t-stat 2.48 with a significance PLATEAU across the grid (t 2.0-2.7), spike-robust (74% of PnL survives removing the top 2 winners), and an interior optimum on every axis. But these are offline-replay figures; the sandbox is a single year (Sharpe 0.821, PSR 0.826) whose sharpe_ci_low is still -0.809 (CI straddles zero on 37 trades). The analyst/optimizer must confirm the real full-history Deflated-Sharpe clears the floor and the walk-forward OOS holds — a t-stat of 2.48 is a genuine candidate, not a lock, against a 0/213 class prior.

Confirm the 60% notional cap and 2% per-trade risk are acceptable for a single-name always-in book at the portfolio level.

Risk-profile note for the Risk Officer: avg_position_pct is ~51% of equity (band-risk sizing at 2% risk across a wide ATR-band stop can approach the 60% notional cap), max_drawdown 10.0% with CI high 23.9%, and a single sandbox win ($9724) is a large fraction of PnL despite the full-history spike-robustness. The always-in stop-and-reverse structure means near-continuous exposure.

BtcAtrChannelSuperTrendRegimeFiltere-54b520c156

This strategy proposed a volatility-adaptive ATR-channel trend system on BTCUSDT USD-M perps, deliberately distinct from the factory's momentum-confluence survivor, with only three wide-plateau parameters and a slow-EMA regime filter to suppress whipsaw. Its initial 2H backtest over the full 2019-2026 history looked healthy: 320 trades, +482% total return, Sharpe 0.95, profit factor 1.54, +1.20% average net per trade, and a 15.7% max drawdown. However, it never reached optimization or analyst evaluation — the pre-optimization backtest-review stage stalled and timed out three times, leaving no verdict. After 2 iterations it was abandoned for that stuck-review reason, so its edge was neither confirmed nor refuted by the later robustness gates.

A strong initial backtest can be lost to an infrastructure/timeout failure at the review gate rather than a substantive rejection; making the backtest-review step resilient to timeouts would let candidates like this be judged on their actual results.

It was abandoned at the pre-optimization BACKTEST_REVIEW gate, not on merit: the analyst review timed out three times (abandon_reason 'backtest_review_stuck_3_timeouts') with a null verdict, so it never advanced to the 3-phase optimization despite a promising initial backtest.

A long-short, single-instrument, pure-OHLCV trend-following system on BTCUSDT.BINANCE USD-M perps whose direction, entry, and trailing exit are one integrated SuperTrend-style ATR channel, gated by a slow-EMA regime filter to stay out of chop.

On 2H bars over 2019-2026 (28,428 bars, 320 trades) it returned +482.4% total with a Sharpe of 0.95, profit factor 1.54, 41.3% win rate, 15.7% max drawdown, and +1.196% avg net return per trade, comfortably clearing the per-trade viability floor. Optimization was never run, so no walk-forward or holdout robustness numbers exist.