Strategies

BTC USD-M Daily Donchian Breakout, LONG/FLAT Only (Single-Instrument BINANCE Perp, Pure OHLCV, Fixed-Fractional Enter-Once / Hold / Exit-to-Flat with NO Dynamic Resizing and NO Long-Short Flips, Engine-Bug-Avoiding by Construction, Tail-Safe, Low-Parameter)

Outcome: Abandoned

BtcDailyDonchianBreakoutLongFlat

Outcome Summary

BtcDailyDonchianBreakoutLongFlat was engineered as a maximally-clean construction — long/flat only, one fixed-fractional entry, leverage 1.0, pure OHLCV, single instrument — explicitly to avoid the engine's confirmed over-exposure, spot-MTM and aggregation defects, and it backtested cleanly with no explosion and 31 faithful long-only breakout trades. It looked stronger than its abandoned ETH sibling, showing a genuine trend edge (base return +700.2%, profit factor 1.65, alpha 0.125, positive in 5 of 7 years) that earned an 'optimize' verdict. But optimization exposed the edge as 2021-outlier-carried (kurtosis 871, skew 29) and decaying recently, and after multiple-testing correction it failed deflated Sharpe (DSR 0.022, optimized Sharpe 0.472 below the 1.064 luck bar, PBO 0.82) while the untouched time-ordered holdout came in negative (-1.231). On attempt 1, with sensitivity flat except one fragile cliff and the binding failure being a money-losing forward holdout rather than an under-explored region, the analyst abandoned it rather than iterate, matching the pattern of its abandoned BTC Donchian sibling.

Strategy report

Hypothesis

A LONG-ONLY (LONG/FLAT, never short), SINGLE-INSTRUMENT, DAILY Donchian-channel breakout trend strategy on BTCUSDT.BINANCE (USD-M perpetual) deliberately constructed to be BACKTESTABLE under the current engine state, which has three confirmed, strategy-unfixable defects: (1) the USD-M perp base entry pipeline OVER-DEPLOYS exposure to ~2-5x and ignores strategy-level caps/flat-guards — but every proof case shared DYNAMIC daily re-sizing and/or frequent LONG-SHORT FLIPS (the leading diagnosis is a hedging-OMS stacking opposing positions on each flip), while the one recent run that sized CORRECTLY (avg_position_pct 53.7%, metrics_reliable=true) did NOT exhibit that pattern; (2) spot-CASH held-position MTM is broken (so this uses PERP, not spot); (3) multi-instrument equity aggregation is broken (so this is SINGLE-instrument). This design therefore avoids the over-exposure trigger BY CONSTRUCTION: it is LONG/FLAT only (never holds a short, so it can NEVER hold two opposing positions / flip long->short), it uses ONE fixed-fractional position size set at entry with NO daily re-sizing (no vol-targeting, no funding-scaling — the resizing patterns that produced the stacking), and it EXPLICITLY closes to flat before any re-entry. It also restricts signals to OHLCV ONLY (the confirmed deep, wall-free feed; liquidations/OI/account-ratio/taker are all data-walled) and trades the validated daily breakout/trend edge in the with-trend (momentum) direction on BTC, the cleanest core trender. It is distinct from the promoted LS TS-momentum sibling by being long/flat + breakout-entry (a different, simpler return profile) and from the promoted alt Donchian-long books by instrument. Tail-safe: leverage 1.0, hard ATR stop. PRE-REGISTERED ACCEPTANCE: avg_position_pct must be <= ~100% (one position, no stacking) and no single-day return beyond ~+/-15-20% (BTC's max daily move at 1x) — if violated, it is the engine over-exposure defect, not the strategy. Low parameter count.

Implements the hypothesis by construction-avoiding the two confirmed engine over-exposure triggers (dynamic daily re-sizing and long->short flips): should_enter only ever returns BUY (never SELL), so no short is ever held and no flip can stack opposing positions in the hedging OMS; position_size is invoked by the base class ONLY while flat and returns a single equity-fraction notional with no vol/funding scaling, so exposure is one fixed position; the base class exits the full position to flat before any re-entry. Single-instrument (avoids broken multi-instrument equity aggregation) and PERP not spot (avoids broken spot-CASH MTM). Signal is a continuous breakout-depth-in-ATR value that varies every bar (frozen-signal-safe). Tail-safe via leverage 1.0 (notional <= equity, so avg_position_pct <= ~100%) and a hard ATR stop. Daily bars match the long-term breakout/trend edge; low parameter count (8).

Failed both the multiple-testing and forward-holdout gates. Deflated Sharpe DSR=0.0219 (vs 0.95), with the optimized Sharpe 0.472 FAR below the 225-trial expected-max luck bar of 1.064 (is_significant=false, sharpe_ci_low -1.133 deeply through 0, PBO 0.8201 — selection more likely than not overfit). Decisively, the time-ordered HOLDOUT FAILED and is NEGATIVE (holdout_sharpe -1.231, ratio -1.507 vs WF-OOS 0.817): the untouched recent window loses money, reflecting genuine recent decay (2026 -11.98%), and the walk-forward OOS is already carried by a single window ([1.04, 1.49, -0.077], third negative). Sensitivity is flat except a single cliff on donchian_exit (base 0.49 -> 0.80 only at value 8), so the only Sharpe lift is a fragile cliff edge, not a robust region. The edge is outlier-carried (2021 +327%, return_kurtosis 871, skew 29, tail_ratio 0.018 — headline rests on the 2020-21 bull trade); PSR 0.946 vs DSR 0.022 is the classic multiple-testing trap. Not iterate (attempt 1 of 2): sensitivity is flat except one cliff, the best-of-225 Sharpe is far below the luck bar, and the binding failure is a NEGATIVE forward holdout driven by recent-regime decay — not an under-explored parameter region, so a second sweep re-selects another sub-significant, decaying config. Not revise_hypothesis: the trend/breakout mechanism is validated on promoted siblings and BTC is not a dead target — this is single-asset daily BTC breakout being too outlier-dependent and decaying too much recently to clear deflation, matching abandoned BtcCoinMInverseDailyDonchianStopAndReverse (DSR 0.008). FAILURE PATTERN: a clean, low-parameter single-asset BTC daily Donchian long/flat breakout with a better-than-sibling annual spread (5/7 positive years) and a non-overfit IS/OOS ratio STILL fails promotion — its 2021-outlier-carried edge cannot clear the 225-trial luck bar (Sharpe 0.472 vs 1.064, DSR 0.022, PBO 0.82) and its untouched recent holdout is NEGATIVE (-1.231) as the edge decays through 2025-2026. A favorable-looking annual distribution and a non-overfit walk-forward do NOT rescue an edge that fails deflated Sharpe and loses money in the forward holdout.

Implementation

LONG/FLAT-only single-instrument daily Donchian breakout trend strategy on BTCUSDT.BINANCE (USD-M perpetual). Enters long when the daily close breaks above the prior 20-bar high by >0.05 ATR AND is above the 50-bar trend EMA (with-trend momentum). Sizes one fixed-fractional position (90% of equity notional, leverage 1.0) at entry with NO dynamic re-sizing. Exits fully to flat on a hard 3xATR stop or a close below the prior 10-bar Donchian low. Pure OHLCV, never shorts, never flips long->short.

Backtest Review

Sharpe: 0.49
Total return: 700.22%
Max drawdown: 45.91%
Trades: 31
Win rate: 41.9%
Profit factor: 1.65

Clean low-parameter single-instrument OHLCV construction; metrics_reliable=true, plausible ~8x return over 6.5yr (well below BTC buy-hold), no over-exposure explosion (daily_returns are 31 per-trade returns = 31 trades, not impossible portfolio days)

31 long trades is an appropriate sample for a daily 20-bar Donchian breakout; entry_diag shows 31 signaled -> 31 submitted, 0 dropped

Trades faithfully implement the hypothesis: 0 shorts, LONG/FLAT only, uptrend breakout entries

Real, reasonably diversified trend edge: profit_factor 1.65, alpha 0.125, Sortino 5.10, Omega 2.27, positive in 5 of 7 years (only 2022 and 2026 negative) — less outlier-dependent than the ETH sibling

Still outlier-influenced: 2021 +315% dominates, skew 28 / kurtosis 861 / tail_ratio 0.05 — deflated Sharpe & holdout will be the real test post-optimization

Sharpe only 0.49 with sharpe_ci_low -0.77 (CI straddles 0); max_drawdown 46% with ~874-day underwater duration

avg_position_pct 110.7% mildly exceeds the pre-registered ~100% guide — most likely MTM/equity drift during winning holds; optimizer should confirm notional never levers past equity

long/flat only, 0 shorts

long/flat

Analysis

Clean low-parameter construction; walk-forward not overfit by the IS/OOS-ratio test (avg IS 1.14 -> avg OOS 0.82); no over-exposure explosion (avg_position_pct ~95-111% is benign MTM drift)

Trades faithfully implement the hypothesis (27-31 long trades, 0 shorts, breakout entries)

Fails deflated Sharpe decisively: DSR 0.0219 (vs 0.95), optimized Sharpe 0.472 FAR below the 225-trial expected-max luck bar 1.064, is_significant=false, sharpe_ci_low -1.133 (CI through 0), PBO 0.8201

HOLDOUT FAILED and is NEGATIVE: holdout_sharpe -1.231, ratio -1.507 vs WF-OOS 0.817 — the untouched recent window loses money, confirmed by 2026 -12% recent decay

Walk-forward OOS carried by one window ([1.04, 1.49, -0.077], third negative); sensitivity has a cliff on donchian_exit (only lifts at value 8)

Outlier-carried: 2021 +327%, return_kurtosis 871, skew 29, tail_ratio 0.018 — headline rests on the 2020-21 bull trade; PSR 0.946 vs DSR 0.022 is the multiple-testing trap

Outcome Summary

A favorable annual distribution (5/7 positive years) and a non-overfit walk-forward do not rescue an edge that fails deflated Sharpe and loses money in the forward holdout — a 2021-outlier-carried single-asset trend edge that decays through 2025–2026 cannot clear the multiple-testing luck bar.

The analyst abandoned it at the post-optimization ANALYZING stage on attempt 1: it failed deflated Sharpe decisively (DSR 0.0219 vs 0.95, optimized Sharpe 0.472 far below the 225-trial luck bar of 1.064, is_significant=false, sharpe_ci_low -1.133, PBO 0.8201) and — most decisively — the time-ordered holdout was negative (holdout_sharpe -1.231, ratio -1.507), meaning the untouched recent window lost money.

A long/flat-only, single-instrument daily Donchian-channel breakout trend-follower on BTCUSDT.BINANCE perp using pure OHLCV, with one fixed-fractional entry, no shorts, no dynamic resizing and a hard ATR stop at leverage 1.0, deliberately built to be backtestable while sidestepping the engine's known over-exposure, spot-MTM and aggregation defects.

The base backtest returned +700.2% over 31 long trades (0 shorts) with profit factor 1.65, Sharpe 0.49 (CI low -0.77), Sortino 5.10, alpha 0.125, positive in 5 of 7 years and a 45.9% max drawdown; optimization lowered total return to +453.7% with Sharpe 0.47, but the edge stayed concentrated in 2021 (+327%, kurtosis 871, skew 29) with recent decay (2026 -12%).

Backtest and paper results are hypothetical. Trading involves risk of loss.