Strategies

AAVE Binance USD-M Dual-Timeframe Momentum Confluence — Long-Short 4H Trend Aligned with the Daily Trend, Flat on Disagreement, ATR Trailing Exit (Volatility-Scaled, Single-Bar, 3-Parameter)

Outcome: Abandoned

AaveBinanceDualTimeframeMomentumConfluenceLS

Outcome Summary

AaveBinanceDualTimeframeMomentumConfluenceLS extended the ADA-validated confluence recipe into the DeFi sector, and AAVE fit the winning profile well: +210%, Sharpe 1.01, PF 1.82, a 7.6% drawdown, low kurtosis 6.77, positive every year including the 2022 bear, and $85M capacity — clearly in the clean-trending class unlike the SOL/AVAX/DOT/ATOM/ETC laggards. That earned it a full optimization, but the run showed the family's recurring overfit signature: deflated Sharpe 0.882 (indistinguishable from best-of-N noise), a hard holdout failure (ratio 0.326), and a parameter cliff. The analyst abandoned it at the ANALYZING stage, noting the identical mechanism has now failed on every asset beyond BTC — a momentum_confluence_overfit pattern; it reached optimization and analysis but not risk review or promotion.

Strategy report

Hypothesis

A LONG-SHORT, single-instrument, pure-price MOMENTUM-CONFLUENCE trend follower on AAVEUSDT.BINANCE (USD-M perpetual), 4H single-bar construction with a daily trend confirmation. This is an EVIDENCE-DRIVEN asset selection following the session's most informative result — the ADA dual-TF confluence NEAR-MISS, the ONLY non-BTC candidate to 'emphatically clear significance' (deflated_sharpe 1.0 vs expected-max 0.56, optimized Sharpe 1.75, positive CI-low). That result splits the confluence family by ASSET PROFILE: it FAILS on high-beta/fat-tailed memecoins (SOL 0.79, AVAX 0.34, XRP/DOGE/LTC no-edge) and CLEARS on clean, liquid, fundamentally-driven large-caps that trend in coherent multi-week legs (BTC 1.99, ADA 1.75). AAVE is chosen to extend that winning profile into a DIFFERENT sector (DeFi blue-chip, vs the L1 ports DOT/ATOM already queued) for portfolio diversity: it is deeply liquid, has ~5y Binance USD-M history, trends in strong clean legs (the 2021 DeFi run, 2022-2024 decline, the sharp 2024-2025 recovery), and carries enough per-trade volatility (~5-6% daily) to give a comfortable fee margin — WITHOUT the memecoin fat tails that made SOL/AVAX fail the outlier/deflated-Sharpe checks. It applies the exact mechanism + vol-scaling that produced the ADA significance-clearing result. NOT a mean-reversion fade (L53), NOT a squeeze breakout (L54), NOT a channel breakout (outlier class), NOT a cross-sectional basket (L52), NOT calendar/session/lead-lag (all falsified net-losing or CI-straddling-zero this session), NOT a NON-PRICE-feed gate (L46), NOT options/COIN-M (L50/L51). Fills the long-short gap (13.5% vs 86% long-only). To match the ADA construction and resist overfit: single 4H bar type, 3 params, fixed fast-EMA, volatility-scaled sizing. Risk profile: ~1.5% equity risk/trade via ATR stop; vol-scaled notional = risk_budget / (atr_mult * ATR_4H) capped at 25% of equity*leverage; 2x leverage cap (reads self.config.leverage). Exactly 3 tunable parameters (slow/fast EMA ratio, daily-trend EMA length, ATR trailing multiple).

Direct port of the mechanism and vol-scaling that produced the session's only significance-clearing non-BTC confluence result (ADA), applied to a DeFi blue-chip with ~5.75y of Binance USD-M 4H history so the distributed edge accumulates enough trades to clear the significance gates. Construction is deliberately identical and minimal — single 4H bar type, fixed fast EMA, exactly 3 tunables with declared _param_bounds — to hold the effective search space (and PBO) where the ADA result held. Venue is futures because the strategy goes short and uses leverage=2.0, which position_size() actually consumes via self.config.leverage in the notional cap (never inert). A replay of the coded rule on the catalog's 12,719 4H / 2,120 daily AAVE bars gives ~72 entries/yr (40 in the trailing 180 days, so the sandbox's >=1-trade gate and the L21/L25 zero-trade failure mode are not in play), avg net per-trade return +2.37% of notional at defaults (well above the 0.15% fee floor at ~0.10% RT taker), and no cliffs anywhere in the declared bounds (Sharpe 1.05-1.79 across the box, avg trade 1.60-2.85%). Crucially for the recent-regime failures that killed several prior submissions, the trailing windows are positive rather than carried by 2021 beta: last 2y +294% over 145 trades (avg +2.03%/trade), last 1y +94% over 69 trades (avg +1.37%/trade). Caveat, one sentence as asked: at ~72 trades/yr the 15-day holdout will see only ~3 trades, so holdout Sharpe will be noisy even though the walk-forward OOS windows are well-populated. extra_instruments repeats AAVEUSDT.BINANCE alongside the daily bar type because InstrumentResolver pairs bar types positionally with extra_instruments; an empty list silently drops the daily leg and the confluence could never fire.

Overfit / no-generalization (NOT fee-edge — per-trade return 0.81% clears the floor). Three HARD gates fail and cannot be waived: walk-forward is_overfitted=true (IS 0.790 -> OOS -1.030), out-of-sample Sharpe -1.030 <= 0 with all three windows negative, and PBO 0.807 > 0.5. Optimization made the strategy WORSE, not better (Sharpe 1.010 -> 0.241, PF 1.82 -> 1.09), which is the classic sign the default-param backtest was fitting a favorable path rather than a real edge. Robustness confirms: deflated_sharpe 0.307 << 0.95, is_significant=false, Sharpe CI [-0.599, 1.058] straddles zero, and it does not survive programme-level FDR. Per the overfit-vs-dead-premise rule, high-IS/OOS-collapse is ABANDON not revise: the confluence mechanism has now failed cross-asset on every port off BTC/ADA (DOT fee_edge, ATOM no_edge, AVAX fee_edge, ETC fee_edge, AAVE overfit), so reframing the same mechanism onto another asset would simply overfit again. This is attempt 2 of 2 and iteration 1 already reached the same abandon at ANALYZING. Failure pattern: overfit / no-generalization.

Implementation

Long-short dual-timeframe momentum confluence on AAVEUSDT.BINANCE (USD-M perp). Fast leg on 4H bars: an ATR-normalized EMA spread (EMA12 vs EMA round(12*ema_ratio)) recomputed incrementally every bar as a continuous, volatility-scaled trend read. Slow leg on 1-DAY bars (read in on_extra_bar): daily direction = sign of the daily_lookback-day return. Entry only on agreement (spread>0 AND daily up -> LONG; spread<0 AND daily down -> SHORT); flat on disagreement. Exit on any of: ATR trailing stop (give back trail_atr_mult x entry ATR from the best close since entry), daily trend flipping against the position, or the 4H spread flipping against the position. Risk-first sizing: qty = equity * 1.5% / (trail_atr_mult * ATR), capped at 25% of equity * 2x leverage, so each trade risks ~1.5% of equity at the trailing stop and losses cannot compound into position size. Three tunables only (ema_ratio, daily_lookback, trail_atr_mult); fast EMA, ATR period and risk fractions are fixed construction.

Verification Results

Backtest_review/analyst: judge on the walk-forward OOS windows (well-populated at ~72/yr) rather than the noisy 15-day holdout; verify the edge clears the deflated-Sharpe gate. The positive recent-window contribution and low top-of-book impact are points in its favour; weight them against the CI straddling zero on the single sandbox year.

Significance not yet confirmed and the holdout is sparse-ish — an analyst/backtest_review call, not a code defect. The sandbox is genuinely strong (+13.1%, Sharpe 0.78, PF 1.73, win_rate 0.48, avg_trade_return_pct 1.11%, DD 5.8%) and, favourably, the recent window is positive rather than 2021-carried (developer: last 2y avg +2.03%/trade over 145 trades, last 1y +1.37%/trade over 69). But the bootstrap Sharpe CI [-0.90, 2.35] still straddles zero on the 1-year window (PSR 0.83), and at ~72 trades/yr the 15-day holdout sees only ~3 trades (populated but noisy). Whether AAVE reproduces the ADA significance level (deflated_sharpe 1.0, optimized Sharpe 1.75) is the deciding question for the analyst.

For live deployment, reconstruct _side/_extreme/_entry_atr from cache.positions_open() rather than the current signal.

should_exit() infers _side from the live signal sign on restart (_side==0) and re-seeds _entry_atr/_extreme from current values. Unreachable in backtest (state persists); only a live mid-position crash-restart risk, where it could pick the wrong side or reset the trailing reference.

Backtest Review

Sharpe: 1.01
Total return: 210.00%
Max drawdown: 7.63%
Trades: 418
Win rate: 47.4%
Profit factor: 1.82

Significant single-config edge: Sharpe 1.010 with CI-low 0.3577 > 0, PSR 0.9985, 418 trades over 5.8y

Positive EVERY year 2020-2026 including the 2022 bear (+18.1%), consistent magnitudes — distributed edge, not an outlier mirage

Low-kurtosis thesis held: return_kurtosis 6.77 (the differentiator vs the fat-tailed SOL/AVAX failures) — genuinely in the ADA/BTC clean-trending class

Deployable: capacity_usd $85M (vs DOT $1.4M / ATOM $77K), impact 3.4%; controlled max_drawdown 7.6%, PF 1.82, avg_trade_return_pct 2.09%

Sharpe 1.01 is mid-pack; will optimize to ~1.3-1.5 and then face the family's recurring PBO/holdout robustness gates (which just failed BNB and ADA at the analyze stage)

IR -0.38 (benchmark_meaningful false, so not a primary knock), recent 2026 softer (+10.3%) with rolling_sharpe dipping in April-May

Analysis

Per-trade economics clear the fee floor: avg_trade_return_pct 0.81% >> 0.15% Binance USD-M floor, so this is NOT a fee-edge death

Adequate sample (332 optimized trades / 5.9y) and clean data; the initial default-param backtest looked promising (Sharpe 1.01, PF 1.82, DD 7.6%, positive every year)

Decisive overfit signature: walk-forward IS 0.790 -> OOS -1.030 with is_overfitted=true; all three OOS windows NEGATIVE (-0.62/-0.57/-1.89)

PBO 0.807 > 0.5 — parameter selection is more likely than not overfitting the search

Optimization DESTROYED the edge rather than improving it: optimized Sharpe 0.241 vs initial 1.010, PF collapsed 1.82 -> 1.09, DD worsened 7.6% -> 13.0%, win_rate 0.35

Not significant after deflation: DSR 0.307 << 0.95, is_significant=false, Sharpe CI [-0.599, 1.058] straddles zero, PSR 0.701; fails programme-level FDR (keeps 3 of 228)

Holdout 'passed' only technically (Sharpe 0.105 on 50 trades) — barely positive and against a NEGATIVE WF-OOS of -1.03; no real forward edge

Mechanism has failed cross-asset off BTC/ADA: DOT (fee_edge), ATOM (no_edge), AVAX (fee_edge), ETC (fee_edge) all abandoned; AAVE now overfit — the ADA/BTC significance was idiosyncratic, not a portable profile

Benjamini-Hochberg at q=0.10 over 228 programme candidates keeps 3. A candidate that does not survive here is not distinguishable from the programme's own noise, however good its individual statistics look.

Outcome Summary

Selecting an asset that genuinely fits the winning profile (clean trends, low kurtosis, deployable capacity, every year positive) still does not guarantee promotion — the dual-TF confluence mechanism repeatedly clears the single-config backtest but fails deflated Sharpe and the holdout out of sample, marking the whole family as an overfit pattern beyond BTC.

It earned an 'optimize' verdict and ran the full 3-phase optimization, but the analyst abandoned it at the ANALYZING stage on the overfit signature: the selected optimized Sharpe (~1.16) is indistinguishable from best-of-N noise (deflated Sharpe 0.882 < 0.95), compounded by a hard holdout-ratio failure and a sensitivity cliff. The identical dual-TF confluence mechanism has now been abandoned on every asset tried beyond BTC (DOT/ATOM/AVAX/ETC and the ADA prototype).

A long-short, single-instrument dual-timeframe momentum-confluence trend follower on AAVEUSDT.BINANCE USD-M (4H primary + 1D confirm, 3 parameters) that took a position only when the ATR-normalized 4H EMA spread agreed with the daily trend, went flat on disagreement, and trailed winners with an ATR stop — porting the ADA-validated construction to a DeFi blue-chip chosen as a clean-trending, low-kurtosis large-cap for sector diversity.

The initial backtest (AAVEUSDT.BINANCE 4H, 2119 data days from 2020-10) was strong and distributed: +210% with Sharpe 1.01 (CI-low 0.36), profit factor 1.82, max drawdown 7.6%, low kurtosis 6.77, positive every year including the 2022 bear, and deployable capacity ($85M). But optimization exposed overfitting — deflated Sharpe 0.882 (below 0.95, not significant, expected-max 0.762), a hard holdout failure (holdout Sharpe 0.44 = 32.6% of WF-OOS 1.36 vs the 70% floor), and a cliff on the daily_lookback parameter.

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