DualTimeframeConfluenceNetDirectionalMomentumBasketLS

Diversified Dual-Timeframe-Confirmed Net-Directional Momentum Basket Across Majors (Binance USD-M, Daily+Weekly Trend Confluence, Flat-on-Disagreement, Weekly Flatten-and-Reopen)

A NET-DIRECTIONAL, multi-instrument momentum basket across ~10 deep-history Binance USD-M majors (BTC, ETH, SOL, BNB, XRP, ADA, AVAX, LINK, LTC, DOGE) whose DISTINCT differentiator is a DUAL-TIMEFRAME TREND-CONFLUENCE filter: each asset is held LONG only when BOTH its medium-horizon DAILY trend AND its WEEKLY trend agree up, SHORT only when both agree down, and FLAT whenever the two timeframes disagree. This stays strictly inside the ONLY validated lane this session — the diversified net-directional momentum basket (the promoted Absolute Momentum basket passed its holdout; single-asset TSM on BTC/BNB/XRP/LTC and dollar-neutral/vol-managed cross-sectional momentum all failed) — while adding an additive robustness mechanism that directly attacks the whipsaw-at-trend-turns failure mode that destroyed every single-asset variant (the ER-gate could not fix it on BTC 4H; LTC chopped to -92%). The flat-on-disagreement rule reduces turnover and false flips, materially changing the exposure/turnover profile versus a single-timeframe basket, so it is complementary rather than a re-expression. It explicitly AVOIDS every dead/blocked path: NOT single-asset (the confirmed-dead family), NOT dollar-neutral/relative (decayed in the tight-correlation regime), NOT vol-managed or continuous vol-scaling (falsified + engine notional-inflation artifact), NOT reversion/intraday (anti-predictive/decays forward), NOT options (engine block), NOT spot CASH (over-fill bug), NOT COIN-M non-BTC (no instrument defs). It fills the long_short and multi-instrument buckets. Sizing is engine-safe by construction: discrete WEEKLY flatten-and-reopen to equal-weight targets + a per-asset hard notional cap (the sizing-clean pattern that held the promoted basket at ~24% exposure), so there is no continuous rebalance and no winner-balloon. Two core parameters (daily_lookback, weekly_lookback).

Stays strictly in the one validated lane (diversified net-directional momentum basket; the promoted Absolute Momentum basket passed its holdout) while adding an additive robustness mechanism that attacks the whipsaw-at-trend-turns failure that destroyed every single-asset TSM variant. The distinct differentiator is the dual-timeframe confluence per-leg signal: a leg only takes a side when its daily (daily_lookback) and weekly (weekly_lookback*7 days) trend signs AGREE, and stands flat on disagreement -- which de-risks regime transitions and cuts turnover/false flips, materially changing the exposure profile versus a single-timeframe basket (complementary, not a re-expression). It keeps every validated structural choice: EQUAL-WEIGHT (not inverse-vol/vol-managed, which the analyst falsified and which inflates notional), NET-DIRECTIONAL via the signed sum of active legs, and weekly discrete rebalance. Sizing is engine-safe by construction: each active leg targets sign*equity*gross_exposure/N and every rebalance moves each leg to its fresh signed target with a single netting delta order (the bounded flatten-and-reopen end-state, no continuous rebalance, no winner-balloon), with an aggregate gross cap as a hard backstop so total gross never exceeds 90% -- avoiding both the uncappable held-winners pattern and the vol-scaling notional-inflation artifact. All order management is centralized in calculate_signal, fired by the base only after the cross-leg alignment barrier confirms every same-daily-timeframe leg is contemporaneous; the abstract hooks are inert. min_bars_required = max(daily_lookback, weekly_days)+2 and a bar counter gate the first rebalance; the tanh log-return proxy is a continuous per-bar signal so Layer-2 frozen-signal detection passes while trades stay schedule-gated. Two tunables keep the overfit surface minimal. All 10 constituents are 2020-listed liquid majors with full daily history (insufficient-history/price<=0 legs are skipped that rebalance), so no leg risks data_unavailable. leverage stays 1.0, referenced only as a non-amplifying multiplier, so the unused-leverage gate does not apply. Pure OHLCV avoids the tick/options/funding/spot-CASH/COIN-M engine walls the hypothesis enumerates. If the multi-instrument daily-return aggregation artifact seen elsewhere this session recurs, this strategy's gross is correctly capped <=90%, so that is the harness accounting bug to escalate rather than a strategy sizing error.

Overfit with negative forward edge AND failed deflated Sharpe. Walk-forward is_overfitted=TRUE: avg in-sample Sharpe 1.161 collapses to avg OOS -0.136 (windows -2.575/+1.418/+0.748, one deeply negative), so the strategy does not generalize out-of-sample. The multiple-testing gate concurs: deflated_sharpe=0.0001 (vs 0.95 bar), is_significant=false, with the optimized Sharpe 0.451 far BELOW the expected-max best-of-225 luck bar of 1.6199, and sharpe_ci_low=-0.2352 straddling 0. The holdout 'pass' is spurious — ratio is computed as 0 against a NEGATIVE WF-OOS Sharpe (-0.136), so the gate is uninformative. The optimizer chased 2021 mania (annual_returns 2021 +709%) by pushing gross above the design's own caps (best_params max_gross 1.306, gross_exposure 1.157 > 1.0), producing a +1232% headline that is 321.7% un-exited open mark-to-market with a 60.7% max drawdown (WORSE than the base 45.8%) and return_kurtosis 178.8 — paper profit, not a realized risk-adjusted edge. Not iterate (attempt 1 of 2): sensitivity is already flat with 0 cliffs (optimizer on its plateau) yet the walk-forward is overfit with negative average OOS, so there is no robust region to tune toward — a second best-of-225 sweep re-fits the 2021 bull and re-inflates gross. Not revise_hypothesis: the net-directional momentum-basket mechanism is already validated on the promoted Absolute Momentum sibling; this is the daily+weekly confluence VARIANT failing deflation on the same 10-major universe (variant-selectivity), not a proven mechanism stranded on a dead target. FAILURE PATTERN: layering a dual-timeframe (daily+weekly) confluence filter onto the validated net-directional basket fails forward — the filter thins exposure enough that the genuine edge drops below the 225-trial luck bar (1.62), and best-of-N selection then levers gross ABOVE the design caps to recover in-sample 2021 return, yielding an is_overfitted walk-forward with negative average OOS, DSR ~0, a CI through 0, a 60% drawdown, and a headline that is mostly open MTM. This is the second confluence-variant of the session to fail this way; the promoted Absolute/single-timeframe basket is where this mechanism class clears the bar.

Long-short net-directional momentum basket across 10 full-history Binance USD-M majors (BTC, ETH, SOL, BNB, XRP, ADA, AVAX, LINK, LTC, DOGE), daily bars, weekly rebalance. Each leg requires dual-timeframe trend agreement: it goes long only when both the 30-day daily trend and the ~10-week weekly trend are up, short only when both are down, and stays flat when the two timeframes disagree. Weighting is equal-weight (each active leg = gross_exposure/N notional, signed), so the book is net-directional with gross <=90%. Every rebalance moves each leg to its fresh signed target via a single netting delta order with an aggregate gross cap backstop. Pure OHLCV, leverage 1.0. Two tunables: daily_lookback, weekly_lookback.

Real (thin) edge: profit_factor 1.166 (>1.0), CAGR +25.7%, alpha +0.19 — above the abandoned sub-1.0 basket siblings (consensus 0.968, inverse-vol 1.0005, long-horizon 0.96)

Validated diversified net-directional structure (promoted absolute-momentum basket, Sharpe 2.17); flat-on-disagreement confluence is a reasonable whipsaw filter

Two non-mania positive years (2024 +40%, 2026 +40%) plus low fee drag (1.49% of gross) and sane per-leg sizing (avg_position 14.7%)

Near-DUPLICATE of the dual-TF-confluence basket already in optimization (same class; 60d/120d twin had PF 1.211) — this only changes the lookback horizons the optimizer already searches, and is marginally weaker (PF 1.166)

2021-concentrated: +346% (2021 mania) dominates; negative in 4 of 7 years (2020, 2022 -24%, 2023 -33%, 2025 -16%)

Outlier-carried + MTM-inflated: return_kurtosis 157, +66.6% single day (2024-12-24), end_unrealized_pct 170%

Thin PF (1.166), Sharpe 0.425 with CI straddling 0 (sharpe_ci_low -0.26), max_drawdown 46%

3 of 7

majority

no single-leg domination

Clean sensitivity surface: 0 cliffs across all 7 parameters; PBO 0.4471 (<0.5)

Differentiated concept (daily+weekly confluence filter, flat-on-disagreement) and engine-safe discrete weekly flatten-and-reopen sizing structure

Walk-forward is_overfitted=TRUE: avg IS 1.161 collapses to avg OOS -0.136, with one window deeply negative (-2.575) — net-negative out-of-sample generalization

FAILS deflated Sharpe: DSR=0.0001 (vs 0.95), is_significant=false; optimized Sharpe 0.451 is far below the expected-max luck bar 1.6199 over 225 trials

sharpe_ci_low=-0.2352 — 95% CI straddles 0

Holdout 'passed' is spurious: ratio=0 against a negative WF-OOS Sharpe (-0.136); the gate cannot validate when the denominator is negative

Optimizer breached the design's own caps (max_gross 1.306, gross_exposure 1.157 > 1.0), inflating the headline to +1232% return that is 321.7% un-exited open MTM with a 60.7% max drawdown (worse than the base 45.8%) and return_kurtosis 178.8

Edge concentrated in 2021 mania (annual +709%) with negative 2023 (-49%) and 2025 (-11%) — regime-dependent, not a robust forward edge

This was the second dual-timeframe-confluence variant of the session, applying a daily+weekly trend-agreement filter (flat on disagreement) to the promoted net-directional momentum basket across 10 majors. The initial backtest showed a thin but real edge — profit factor 1.166, CAGR +25.7%, alpha +0.19 — marred by a 45.8% drawdown and 2021-concentrated, outlier-carried returns, so it was cleared to optimize. Optimization then exposed it as overfit: the walk-forward's in-sample Sharpe of 1.16 inverted to a negative -0.136 OOS average, deflated Sharpe came in at ~0.0001 against a luck bar of 1.62, and the optimizer breached the strategy's own gross caps to chase 2021 mania, inflating the return to +1232% while deepening the drawdown to 60.7% and leaving 321.7% in un-exited open MTM. The analyst abandoned it on the first iteration as variant-selectivity, noting the promoted absolute/single-timeframe basket is where this mechanism class actually clears the bar.

Layering a dual-timeframe confluence filter onto the validated net-directional basket thins directional exposure enough that the genuine edge drops below the multiple-testing luck bar, after which best-of-N selection levers gross above the design caps to recover 2021 bull-market returns — producing an overfit, negative-OOS result whose headline is mostly open mark-to-market rather than a robust forward edge.

The Strategy Analyst abandoned it at the post-optimization stage: walk-forward flagged is_overfitted=TRUE (avg IS Sharpe 1.161 collapsing to avg OOS -0.136, one window at -2.575), deflated Sharpe was 0.0001 (vs the 0.95 bar) with the optimized Sharpe 0.45 far below the best-of-225 luck bar of 1.62, sharpe_ci_low -0.2352 straddled zero, and the holdout 'pass' was spurious (ratio 0 against a negative WF-OOS Sharpe).

A net-directional momentum basket across 10 deep-history Binance USD-M majors that held each leg long only when its daily and weekly trends both agreed up, short only when both agreed down, and flat on disagreement — adding a dual-timeframe confluence filter to the validated net-directional basket structure to suppress the whipsaw-at-trend-turns failure mode, with weekly equal-weight flatten-and-reopen sizing.

The initial daily-bar backtest over ~2,017 days produced 638 trades, a Sharpe of 0.42 (CI low -0.257), profit factor 1.166, CAGR +25.7% and ~424% total return, but a 45.8% max drawdown with heavily outlier-carried, 2021-concentrated returns (kurtosis 157). After optimization the headline ballooned to +1232% return and Sharpe 0.45, but only because the optimizer breached the design's own caps (max_gross 1.306, gross_exposure 1.157 > 1.0), and the result carried a 60.7% max drawdown with 321.7% of the total in un-exited open mark-to-market.