Strategies

BTC Daily Multi-Week Trend Continuation Long-Only (BINANCE USD-M Futures, Daily, OHLCV-Only)

Outcome: Abandoned

BtcDailyMultiWeekTrendContinuationLong

Outcome Summary

This strategy was the BTC leg of a multi-week trend-continuation triad, directly analogous to the already-landed LINK version and a recently-submitted ETH sibling, using a simple 3-parameter breakout-plus-SMA entry and an ATR trailing exit on pure daily OHLCV. Its base backtest passed the pre-optimization gate on headline strength (862.7% return, profit factor 2.20, alpha +0.169), but reviewers flagged severe outlier dependence — return kurtosis 431.9, a 2021 that returned +395%, and ~189.5% of the total sitting in an open MTM position. Optimization confirmed the concern: the best-of-225 config carried a deflated Sharpe of just 0.0512 with its Sharpe below the multiple-testing luck bar, a PBO of 0.83, a failed holdout (ratio 0.403), and a negative information ratio against buy-hold, with the refit flat-to-negative every year after 2021. After two iterations the analyst abandoned it, recording that inheriting a proven mechanism does not confer forward-generalizing edge on a different instrument.

Strategy report

Hypothesis

Long-only BTC daily strategy on BINANCE USD-M futures. Direct analog of the already-landed LinkDailyMultiWeekTrendContinuationLong mechanism (and the recently-submitted ETH version), applied to BTC as the third instrument in the multi-week trend-continuation triad. Enter when BTC is in a well-established multi-week uptrend confirmed by a fresh higher high. Single-instrument, OHLCV-only, single-condition entry, factory-approved instrument and timeframe (BTCUSDT.BINANCE 1D).

Iteration 2 fix: the reported failure was pyramiding — every fresh higher-high breakout while already long submitted another 1.2x tranche, driving average exposure to ~1.9x and producing impossible single-day return figures. The smallest correct fix is an open-position guard at the top of should_enter(): `if self.cache.positions_open(instrument_id=self.instrument_id): return None` before the _armed check. This enforces one-position-per-signal so cumulative notional stays at the stated 1.2x, bounds daily returns, and lets the ATR-Chandelier/SMA exits fully close positions (reducing the large end-of-backtest unrealized stack). All signal logic, imports, sizing, and exit logic that already passed Layers 1–3 are unchanged — only the entry guard was added, so no earlier layer regresses.

Fails deflated Sharpe, the forward holdout, and multiple-testing robustness on optimization attempt 1. DSR=0.0512 (vs 0.95 bar) with the optimized Sharpe 0.856 BELOW the 225-trial expected-max luck bar of 1.723 (is_significant=false, sharpe_ci_low -0.87 straddling 0, PBO 0.8267 >>0.5) — after multiple-testing correction the selected best-of-225 config is statistically indistinguishable from best-of-N noise; probabilistic_sharpe 0.9466 vs DSR 0.0512 is the textbook PSR-vs-DSR selection-inflation trap. The time-ordered HOLDOUT FAILED (ratio 0.403 < 0.70; holdout_sharpe 0.342 vs WF-OOS 0.849) with OOS window 3 already negative (-0.21). The strategy carries a NEGATIVE information_ratio in both base (-0.155) and optimized (-0.55) forms — it loses to its own BTC buy-hold risk-adjusted — and its edge is bull-run-concentrated: base annual 2021 +395% with a single +205% day (kurtosis 431.9), and the optimized refit is flat/negative post-2021 (2022 -2.2%, 2024 +0.7%, 2025 -7.6%, 2026 +0.9%). The clean 0-cliff sensitivity surface and the non-overfit-flagged walk-forward measure consistency, not significance or forward-generalization, and do not rescue a config that fails deflation and the holdout. Not iterate: the binding failures are structural (best-of-225 Sharpe below the luck bar, failed holdout, PBO 0.83, 2020-2021-concentrated edge with negative alpha since), leaving no robust region above significance to tune toward — a second sweep re-overfits the same bull run. Not revise_hypothesis: BTC is a fine instrument and this mechanism's LINK sibling is already landed, so the mechanism is not stranded on a dead target — the BTC port simply does not generalize forward, and abandoning loses no validated edge. FAILURE PATTERN: transferring the promoted LINK multi-week trend-continuation mechanism to single-asset BTC does NOT inherit promotability — the BTC port produces an in-sample-attractive refit (clean 0-cliff sensitivity, non-overfit walk-forward, Sharpe 0.856) that collapses on the decisive gates: DSR 0.0512 with the optimized Sharpe below the 1.723 luck bar, PBO 0.83, a failed holdout (ratio 0.403), a negative information ratio vs buy-hold, and an edge concentrated entirely in the 2020-2021 bull run (kurtosis 432, +205% single day).

Implementation

Long-only BTC daily trend-continuation on BINANCE USD-M futures. Enters long when the close breaks above the prior 20-day high AND is above the 50-day SMA (multi-week uptrend + fresh higher high). Exits on a Chandelier ATR trailing stop (peak minus 3xATR) or a close back below the SMA. Pure OHLCV, single-instrument, three core parameters. Leverage 2.0 with position_pct 0.60 gives ~1.2x equity notional per position, and only ONE position is held at a time.

Backtest Review

Sharpe: 0.70
Total return: 862.67%
Max drawdown: 26.90%
Trades: 62
Win rate: 58.1%
Profit factor: 2.20

Trades match the hypothesis exactly: long-only (62 long / 0 short), breakout+SMA trend-continuation, ~10 trades/year within the stated 10-20/year band.

Genuine positive expectancy at base: profit_factor 2.20, win_rate 0.58, Sortino 7.92, alpha +0.169, CAGR 41.6%.

Clean execution — pure BTC 1D OHLCV, no data-substitution/supp-data risk, one-position guard prevents pyramiding, 62 entries all submitted (no drops).

Direct analog of the already-landed LINK multi-week trend-continuation mechanism; sample size is adequate for meaningful optimization.

Severe outlier dependence: return_skew 19.6, return_kurtosis 431.9, and annual 2021 +395% carries nearly the whole result while 2022/2025 are negative — the edge may be one-bull-run deep.

Sharpe is only 0.696 with CI straddling zero (sharpe_ci_low -1.13) — significance not yet established; DSR/PBO robustness gate will be decisive.

end_unrealized_pct 189.5 — much of the 862% headline is an un-exited open position (MTM), realized track record is far smaller.

max_drawdown 26.9% with an 788-day underwater duration.

Analysis

Clean mechanics: pure BTC 1D OHLCV, long-only (62/89 trades all long), faithful to the 3-parameter hypothesis, no data-substitution risk.

Sensitivity PASSED with 0 cliffs and the walk-forward is not overfit-flagged (avg IS 1.315 -> avg OOS 0.849).

The underlying mechanism is proven — its LINK sibling is already landed.

Deflated Sharpe FAILS: DSR 0.0512 (vs 0.95), is_significant=false, sharpe_ci_low -0.87 (straddles 0); optimized Sharpe 0.856 is BELOW the 225-trial expected-max luck bar of 1.723.

PBO 0.8267 (>>0.5) — parameter selection is overwhelmingly likely overfitting; probabilistic_sharpe 0.9466 vs DSR 0.0512 is the selection-inflation trap.

Holdout FAILED: ratio 0.403 (<0.70), holdout_sharpe 0.342 vs WF-OOS 0.849; OOS window 3 already negative (-0.21).

NEGATIVE information_ratio in both base (-0.155) and optimized (-0.55) vs its own BTC buy-hold benchmark.

Edge is 2020-2021 bull-run concentrated: base kurtosis 431.9 with a +205% single day (2021-01-27); optimized refit is flat/negative post-2021 (2022 -2.2%, 2024 +0.7%, 2025 -7.6%, 2026 +0.9%).

Base end_unrealized_pct 189.5 — most of the 862% headline was an un-exited open MTM position.

Do NOT optimize until the sizing matches the stated risk model. SPECIFIC FIX: add an open-position guard at the top of should_enter() — `if self.cache.positions_open(instrument_id=self.instrument_id): return None` — before checking self._armed. Currently every fresh higher-high breakout while already long submits another 1.2x tranche, pyramiding to ~1.9x average exposure (avg_position_pct 189% vs the hypothesis's stated 1.2x) and producing impossible single-day moves (+175.7% on 2021-01-27, +62.2% on 2024-12-27). If pyramiding into trends is actually intended, instead cap cumulative accumulated notional (e.g. max 1.2x total across all tranches) and state it explicitly — but the cleaner path is one position per signal. Separately, the headline 862.7% is ~45% unrealized (end_unrealized_pct 385%): a large stacked long is still open at backtest end because the ATR-Chandelier/SMA exits didn't trigger — after the flat-check fix, re-run and confirm (a) avg_position_pct sits near 120%, (b) daily returns are bounded sanely, and (c) the realized (closed-trade) return and Sharpe still show a positive edge. The underlying trend-following alpha looks real (alpha +0.166, PF 1.76), so this is a sizing-correctness fix, not an abandon — once the clean re-run confirms the edge on realized terms, proceed to optimization.

Outcome Summary

A promoted mechanism does not transfer its promotability to a new instrument — the BTC port produced an in-sample-attractive refit (clean 0-cliff sensitivity, non-overfit walk-forward) whose edge was concentrated almost entirely in the 2020-2021 bull run and collapsed on deflation, holdout, and benchmark-relative gates.

The analyst abandoned it at the post-optimization ANALYZING stage on the robustness gates: deflated Sharpe 0.0512 (vs 0.95) with the optimized Sharpe 0.856 below the 225-trial luck bar of 1.723, PBO 0.8267, a failed time-ordered holdout (ratio 0.403, holdout Sharpe 0.342), and negative information ratio versus BTC buy-hold in both base (-0.155) and optimized (-0.55) forms.

Port the already-landed LINK multi-week trend-continuation mechanism to BTC — a long-only, pure-OHLCV daily strategy on BTCUSDT.BINANCE USD-M futures that buys when price breaks above the prior 20-day high while above its 50-day SMA, then rides the trend with a Chandelier ATR trailing stop.

The base backtest looked strong on the surface — 862.7% total return, CAGR 41.6%, profit factor 2.20, 62 long trades (58% win rate) — but Sharpe was only 0.696 with a CI straddling zero, and ~189.5% of the headline was an un-exited open MTM position. Optimization cut it to a Sharpe of 0.856 and +138.3% return, still with negative information ratio and heavy bull-run concentration.

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