Strategies

XRP 4H Chandelier Exit Trend-Following Long (22, 3.0)

Outcome: Abandoned

XrpFourHourChandelierExitTrendFollowingLong

Outcome Summary

XrpFourHourChandelierExitTrendFollowingLong was a disciplined attempt to ride the session's one approved mechanism class — Tier-1 path-dependent ATR trend filters — by porting LeBeau's Chandelier Exit onto XRP 4H with canonical defaults. Its trend payoff was genuine (avg_win 2.2× avg_loss, +78% raw), and it cleared the review gate, but it carried a fatal fingerprint: return_kurtosis 94 that optimization pushed to 292, meaning the gains lived in a few enormous pump days rather than a persistent edge. The forward gates rejected it decisively — DSR 0.01 against a 1.30 luck bar, a holdout Sharpe of −1.382, an overfitted walk-forward, and deep recent decay (2025 −12.0%) that directly contradicted the positive-2025/2026 Supertrend performance the hypothesis had cited as its rationale. The analyst abandoned it on its second iteration: the viable Tier-1 trend edge is already captured by the promoted Supertrend sibling, leaving XRP-Chandelier as the decaying, fat-tail arm of the class — and joining DOT, SOL, and ADA-momentum as confirmation that weak single-major trend/momentum edges optimize to a flattering Sharpe and then fail honest deflation and a one-shot holdout.

Strategy report

Hypothesis

A long-only single-instrument PATH-DEPENDENT TREND-FOLLOWING strategy on XRPUSDT perpetual futures (Binance USD-M) using the Chandelier Exit (Chuck LeBeau) on 4-hour bars. CRITICAL DESIGN CHOICES driven by the now-definitive session feedback that classifies indicators into 3 tiers: TIER 1 (path-dependent ATR-based filters like Supertrend, Chandelier, Parabolic SAR) — Supertrend approved this session with Sharpe +1.17, PF 1.21, POSITIVE 2025/2026 performance. TIER 2 (channel/volatility-band breakouts: Donchian, Keltner, WRB) — all abandoned for 2025+ regime decay. TIER 3 (oscillator-style entries: MACD, HMA-slope, Aroon) — all abandoned for structural no-edge regardless of regime. This proposal stays strictly in TIER 1: Chandelier Exit uses the extreme high (or low) of the last N bars minus a multiple of ATR — a path-dependent trailing stop that flips discretely when price crosses it, structurally filtering chop just like Supertrend. Distinct from Supertrend (which anchors to median price ± ATR) and Parabolic SAR (which uses acceleration factor on extreme point): Chandelier anchors to the rolling N-bar HIGH minus k×ATR, giving slower, more selective signals. ADDITIONAL DESIGN CHOICES: (a) XRP is NOT in the current portfolio as an OHLCV strategy (the failed XRP entry was the premium_index supp-data hypothesis); XRP has 5+ years of Binance perp data (well above the 'minimum 2 years' rule from session feedback). (b) Long-only on alt-perp is acceptable BECAUSE the Tier 1 mechanism class has demonstrated positive 2025-2026 performance (Supertrend +4.5%/+1.7%). (c) MAXIMALLY SIMPLE — 2 canonical LeBeau parameters (lookback=22, ATR multiplier=3.0) FIXED at defaults.

Iteration 2 fix for the Layer-2 frozen-signal failure. Root cause: the proxy returned (close - ref)/atr clipped to [-1, 1]; in a steady trend the close holds a roughly fixed ATR-distance beyond the trailing stop, so it saturated at exactly +1.0 (uptrend) / -1.0 (downtrend) on all 477 bars. The smallest fix removes the clip (there is no [-1,1] requirement on calculate_signal) and adds a close/ATR level term scaled by direction, which strictly increases in an uptrend (rising close, ~constant ATR) and strictly decreases in a downtrend, guaranteeing the signal varies every bar even in a perfectly linear synthetic trend. Only the return expression in calculate_signal changed; the Chandelier state machine, should_enter/should_exit (which read the discrete _dir, not the signal), position_size, imports, and config are untouched, so Layer-1 static and the already-passing trend logic remain intact and trading behavior is unchanged.

Fat-tail-carried weak trend edge that fails multiple-testing deflation and the forward holdout -- exactly the outcome flagged at pre-optimization. The optimizer improved PF (1.40) and passed sensitivity, but every decisive gate fails: deflated_sharpe 0.01 (is_significant FALSE) with the optimized Sharpe 0.53 BELOW the expected-max luck bar 1.30 over 225 trials, PBO 0.5728, sharpe_ci_low -0.36; the HOLDOUT FAILED catastrophically (holdout_sharpe -1.382, ratio -5.469 -- the recent window loses money badly); and walk-forward is_overfitted=TRUE (avg IS 0.568 -> avg OOS 0.253, one window -0.81). The root cause is fat-tail dependence that the optimizer WORSENED: return_kurtosis rose from 94 to 292, so the +95% headline is carried by a handful of huge XRP-pump days (largest_win $47.8k, +32.6% on 2024-11-24), and DSR's skew/kurtosis adjustment correctly discounts it to ~0.01; the edge is also decaying (2025 -12.0%, 2026 -2.8%, rolling Sharpe ~-6 to -7 through 2026). Sensitivity passing measures only in-sample smoothness and does not rescue a failed holdout + DSR. Not iterate (attempt 1 of 2): the failure is a fat-tail-driven sub-luck-bar edge that won't deflate + a negative holdout + recent decay, not a tunable parameter region -- a second sweep re-finds the same outlier-carried config. Not revise_hypothesis: the viable Tier-1 path-dependent trend expression is already captured by the approved Supertrend sibling (which had POSITIVE 2025/2026, unlike this), so XRP-Chandelier is the decaying/fat-tail arm with no reframe. FAILURE PATTERN: single-asset long-only trend-follower on a crypto major whose headline is carried by a few fat-tail pump days fails deflated Sharpe (0.01, below the luck bar) with a deeply negative holdout (-1.38) and is_overfitted=TRUE -- the fourth weak-trend/momentum strategy this run (after DOT, SOL pivot, ADA momentum) to optimize to a ~0.5-1.2 Sharpe and then fail DSR + a negative holdout. A high raw return with kurtosis ~300 and a 0-straddling CI is outlier luck, not a deflatable edge.

Implementation

Long-only single-instrument Tier-1 path-dependent trend-follower on XRPUSDT.BINANCE USD-M futures 4H bars, OHLCV-only, using Chuck LeBeau's Chandelier Exit (lookback=22, ATR multiplier=3.0). It maintains a discrete trend state from a ratcheting ATR trailing stop: long_stop = highest_high(22) - 3*ATR (ratchets up), short_stop = lowest_low(22) + 3*ATR (ratchets down); the trend flips UP when close crosses above the prior short_stop and DOWN when close crosses below the prior long_stop. It holds long while the trend is up and exits the moment it flips down, re-entering on the next up-flip. Position size risks 1.5% of equity over the distance to the long_stop (capped at 18% notional). calculate_signal returns a continuous, bar-varying ATR-normalized proxy; entry/exit read the stored discrete trend direction.

Verification Results

Verification failed (Layer 2 — synthetic scenarios): Parameters used: ['atr_mult', 'lookback', 'risk_pct', 'atr_period', 'min_notional', 'min_stop_pct', 'per_trade_pct'] Check that __init__ sets all attributes from self.parameters.get(). - steady_uptrend: Frozen signal: all 477 signals are identical (value=1.0). calculate_signal() likely has a bug — the signal never varies. - steady_downtrend: Frozen signal: all 477 signals are identical (value=-1.0). calculate_signal() likely has a bug — the signal never varies.

Backtest Review

Sharpe: 0.42
Total return: 77.97%
Max drawdown: 21.60%
Trades: 555
Win rate: 34.8%
Profit factor: 1.18

Correct trend payoff: win_rate 0.348 with avg_win $1,765 vs avg_loss $797 (2.2x); PF 1.18, total_return +78%, max_dd 21.6%; ample sample (555 trades)

Validated Tier-1 path-dependent ATR mechanism class (Supertrend sibling approved this session); positive in 4 of 7 years (2020/2021/2023/2024)

Extreme fat-tail dependence: return_kurtosis 94.55, skew 8.19 -- headline carried by a handful of huge XRP-pump wins; DSR will heavily discount this

Sharpe 0.42 with sharpe_ci_low -0.42 (CI straddles 0) -- weak, not yet significant; primary DSR-fail risk after selection

Recent decay: 2025 -5.2%, 2026 -4.7% (rolling Sharpe deeply negative) -- opposite of the approved Supertrend's positive 2025/2026, undercutting the hypothesis's own rationale

~0.5+ (pre-opt)

~62 WF floor

4 of 7

majority-ish

moderate (<~10)

2.2x (avg_win/avg_loss)

> 1 for trend

Analysis

Sensitivity PASSED (1 cliff on atr_mult); optimizer lifted PF to 1.40 and cut impact to 3.7%; positive in 4 of 7 years in the optimized run

Functioning Tier-1 trend mechanism (Chandelier), market-neutral-ish beta 0.10

Fails deflated Sharpe decisively: DSR 0.01 (<<0.95), is_significant FALSE; optimized Sharpe 0.53 BELOW the expected-max luck bar 1.30 over 225 trials; PBO 0.5728; sharpe_ci_low -0.36 (straddles 0)

Holdout FAILED catastrophically: holdout_sharpe -1.382, ratio -5.469 -- the recent window loses money badly

Walk-forward is_overfitted TRUE (avg IS 0.568 -> avg OOS 0.253, one window -0.81); return_kurtosis 292 (UP from 94) -- the edge is even more carried by a few huge XRP-pump days

Recent decay: 2025 -12.0%, 2026 -2.8%, rolling Sharpe ~-6 to -7 through 2026

> 1.30 (expected-max)

Outcome Summary

A single-asset long-only trend-follower whose headline is carried by a few fat-tail pump days (kurtosis ~300) fails deflated Sharpe (0.01, below the luck bar) with a deeply negative holdout (−1.38) and is overfitted — the fourth weak-trend/momentum strategy this run (after DOT, the SOL pivot, ADA momentum) to optimize to a ~0.5–1.2 Sharpe then fail DSR + holdout — so a high raw return with a zero-straddling CI is outlier luck, not a deflatable edge, and passing sensitivity doesn't rescue a failed forward holdout.

The backtest-review gate passed it to optimization ('optimize'), but the analyst abandoned it post-optimization: the optimized Sharpe 0.53 sat below the expected-max luck bar of 1.30 over 225 trials (DSR 0.01, is_significant=false), the once-tested holdout failed catastrophically (−1.382), and the walk-forward was overfitted — with the root cause being fat-tail dependence the optimizer worsened (kurtosis 94→292), so the +95% optimized headline is carried by a handful of huge XRP-pump days that DSR correctly discounts, on a decaying edge.

A long-only, OHLCV-only Tier-1 path-dependent trend-follower on XRPUSDT.BINANCE 4H — Chuck LeBeau's Chandelier Exit (lookback 22, ATR multiplier 3.0): a ratcheting ATR trailing stop anchored to the rolling N-bar extreme that flips trend state discretely, holding long while the trend is up and exiting the moment it flips down — deliberately staying in the session's approved Tier-1 mechanism class (alongside the promoted Supertrend) for chop-filtering.

The raw backtest showed a correct trend payoff — 555 trades, win_rate 0.348 but avg_win $1,765 vs avg_loss $797 (2.2×), profit_factor 1.18, total_return +78%, max drawdown 21.6%, positive in 4 of 7 years — but with extreme fat-tail dependence (return_kurtosis 94.55, skew 8.19) and a weak Sharpe 0.42 (CI straddling zero, −0.42 to 1.07). Optimization lifted PF to 1.40 and cut impact to 3.7%, but every decisive gate failed: DSR 0.01, optimized Sharpe 0.53 below the 1.30 luck bar, PBO 0.5728, holdout Sharpe −1.382 (ratio −5.469), walk-forward is_overfitted=TRUE (IS 0.568 → OOS 0.253), and kurtosis rose to 292 with 2025 −12.0% / 2026 −2.8%.

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