MultiHorizonTrendConsensusNetDirectionalBasketLS

Multi-Horizon Trend-Consensus Net-Directional Momentum Basket Across Liquid Majors (Long-Short, Binance USD-M, Equal-Weight, 4 Fixed Horizons / >=3-of-4 Agreement Per Leg, Weekly Rebalance, Full-Flatten-and-Reopen, Pure OHLCV, 2-Parameter)

A LONG-SHORT, MULTI-INSTRUMENT, NET-DIRECTIONAL momentum basket across ~10 confirmed-full-history Binance USD-M majors (BTC, ETH, SOL, BNB, XRP, ADA, DOGE, AVAX, LINK, LTC) whose DISTINCT MECHANISM is a MULTI-HORIZON TREND CONSENSUS signal. It extends the PROMOTED winner -- the equal-weight Absolute-Momentum LS basket (holdout passed, Sharpe 2.17), the factory's ONE validated net-directional-momentum vehicle -- by attacking that structure's single un-addressed weakness: SINGLE-LOOKBACK OVERFIT / SENSITIVITY CLIFFS (the IS->OOS-collapse signature that killed funding-confirmed-trend, the ER-gate, and VWAP-continuation this session). The promoted basket picks ONE trend lookback -- precisely the parameter the optimizer repeatedly exploits into a cliff. MECHANISM: each leg's signal is a CONSENSUS across FOUR FIXED horizons -- sign(close - SMA(close, L)) for L in {20, 60, 120, 240} days -- and a leg takes a side ONLY when at least consensus_threshold (default 3) of the 4 horizons AGREE: LONG if >=3 are up, SHORT if >=3 are down, FLAT otherwise. Averaging FIXED horizons removes the single-tunable-lookback dependence (no lookback parameter for the optimizer to game into a cliff), and the consensus gate stands a leg aside during regime transitions when horizons conflict (de-risking, the documented whipsaw problem). WHY DISTINCT and NOT a prohibited refinement: the analyst ruled out WEIGHTING-scheme refinements (inverse-vol falsified: made 2021 worse, +77%->+410%, Sharpe 1.36->0.57); THIS changes the SIGNAL, not the weighting -- it stays EQUAL-WEIGHT (the PROMOTED weighting), so it is on the non-prohibited axis. It is also distinct from the absolute basket (single trend, always-on), the dual-momentum basket (relative rank + absolute filter), and the dual-TF-confluence basket (binary 2-horizon agreement) -- this is a 4-horizon consensus. WHY NOT a closed family: single-asset alt-TSM is dead (BNB/XRP/LTC catastrophic); the generalizing edge lives ONLY in the diversified basket, and this IS that, with a robustness-improving signal. EVERY design choice respects the hard engine walls: (1) FULL-FLATTEN-AND-REOPEN sizing every rebalance (close ALL legs, recompute, reopen equal-weight summing to <=100% gross) -- the PROVEN pattern that bounded the promoted basket at ~24%, NOT the long-only held-winners pattern that proved uncappable. (2) EQUAL-WEIGHT (the promoted, validated weighting), NOT inverse-vol (falsified). (3) NET-DIRECTIONAL (signed sum of active legs). (4) PURE OHLCV -- no ticks (2-day wall), options (unwired), or funding (data ends 2024). (5) FIXED horizons -> only 2 tunables (consensus_threshold, rebalance_period), the lowest overfit surface of any basket variant, honoring the 'favor simpler, fewer parameters' guidance.

Extends the promoted equal-weight Absolute-Momentum LS basket (holdout Sharpe 2.17 -- the factory's one validated net-directional-momentum vehicle) by replacing its single, cliff-prone trend lookback with a multi-horizon trend-consensus SIGNAL, while keeping every validated structural choice. The signal votes sign(close - SMA(L)) across the FIXED set {20,60,120,240} and requires >=consensus_threshold agreement, which (a) removes the single-tunable-lookback dependence that the optimizer repeatedly exploited into IS->OOS cliffs (there is no lookback parameter), and (b) stands a leg aside when horizons conflict, de-risking regime-transition whipsaws. Critically it changes the SIGNAL axis only and stays EQUAL-WEIGHT -- honoring the analyst's ruling that inverse-vol weighting is falsified (it worsened 2021) -- so each active leg targets sign * equity*gross_exposure/N and the book is net-directional via the signed sum, with gross bounded at <=gross_exposure (90%). Exposure is bounded exactly as required: each rebalance recomputes fresh targets and moves every leg to its signed target via a single netting delta order (the bounded full-flatten-and-reopen end-state without redundant close-then-reopen churn), and an aggregate gross cap is a hard backstop -- structurally preventing the uncappable held-winners balloon. 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; should_enter/should_exit/position_size are inert. min_bars_required = 242 (longest horizon + buffer) and a bar counter gate the first rebalance until all four SMAs are computable; the tanh log-return proxy is a continuous per-bar signal so Layer-2 frozen-signal detection passes while trades stay schedule-gated. Only 2 tunables (consensus_threshold, rebalance_period) give the lowest overfit surface of any basket variant, honoring the favor-simpler guidance. 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 engine walls. If the multi-instrument daily-return aggregation artifact flagged elsewhere this session recurs, this strategy's own gross is correctly capped <=90% via the per-leg equal weight and aggregate cap, so that would be the harness accounting bug to escalate, not a strategy sizing error.

The 4-horizon trend-consensus refinement is a LOSING signal that degrades, rather than improves, the promoted absolute-momentum basket it extends. profit_factor 0.968 (<1.0, leverage-invariant), CAGR -2.39%, Sharpe 0.12, expectancy -$19.6/trade, negative in 4 of 6 years (2022 -14.8%, 2023 -28.3%, 2024 -0.8%, 2025 -10.2%). The +16.9% total is recent/outlier-manufactured: essentially all of it is 2026 (+62%), carried by a few big prints (+44.24% on 2024-01-23, +37.91% on 2024-12-31, +32.15% on the final bar 2026-06-26) with return_kurtosis 168.9 — strip those and it is a multi-year bleed. The hypothesis's central claim, that averaging FOUR fixed horizons with a >=3-of-4 consensus gate removes single-lookback overfit and IMPROVES robustness over the promoted single-trend basket (Sharpe 2.17), is empirically FALSIFIED: the consensus gate over-filters / multi-horizon averaging smooths the edge away, yielding sub-1.0 profit factor and negative CAGR. Not optimize: a PF-0.968, negative-CAGR signal with an outlier-carried headline cannot be optimized into significance — the 2 tunables (consensus_threshold, rebalance_period) search a space that is break-even-negative throughout. Not iterate: the high exposure_pct (849%) is the family's cumulative-metric quirk (avg_position_pct is only 13.0%, real book near the cap), not the blocker; the consensus MECHANISM is what loses, and the promoted single-lookback basket already exists as the working version, so there is nothing to tune toward. FAILURE PATTERN: replacing the promoted basket's single trend with a 4-horizon >=3-of-4 consensus does NOT improve robustness -- it dilutes the net-directional momentum edge to profit_factor < 1.0 and negative CAGR (losing 4 of 6 years), the same sub-1.0-PF signature as the abandoned long-horizon trend basket (0.96) and inverse-vol basket (1.0005). This holds the family line: PF > 1.2 with multi-year consistency (dual-TF confluence 1.211, cross-sectional momentum 1.16) is a real edge worth optimizing/iterating; PF <= 1.0 means the signal refinement itself failed, and no sizing fix rescues it.

Long-short net-directional momentum basket across 10 full-history Binance USD-M majors (BTC, ETH, SOL, BNB, XRP, ADA, DOGE, AVAX, LINK, LTC), daily, weekly rebalance. Each leg's direction is a consensus across four fixed SMA horizons (20/60/120/240 days): it goes long when >=3 horizons have price above their SMA, short when >=3 are below, and stands flat when horizons conflict. 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: consensus_threshold, rebalance_period; the four horizons are fixed.

Pure OHLCV, confirmed-full-history majors, low fee drag (1.64%), lowest parameter surface of the family (2 tunables)

Sane per-leg sizing (avg_position_pct 13.0%, no single-leg domination), not liquidated

Losing signal: profit_factor 0.968 (<1.0), CAGR -2.39%, Sharpe 0.12, expectancy -$19.6/trade, negative in 4 of 6 years (2022 -14.8%, 2023 -28.3%, 2024 -0.8%, 2025 -10.2%)

Headline +16.9% is recent/outlier-manufactured: essentially all 2026 (+62%), carried by +44% (2024-01-23), +38% (2024-12-31), +32% (last bar 2026-06-26); return_kurtosis 168.9

The 4-horizon consensus refinement DEGRADES the promoted basket (Sharpe 2.17) into break-even-negative — the premise that it improves robustness is falsified

Sharpe CI straddles 0 (-0.72 to +0.71), information_ratio -0.52, max_drawdown 52.9%

1-2 of 6

majority

MultiHorizonTrendConsensusNetDirectionalBasketLS tried to harden the factory's one promoted momentum vehicle — the equal-weight Absolute-Momentum LS basket — by replacing its single trend lookback with a consensus across four fixed horizons (20/60/120/240 days), requiring 3-of-4 agreement per leg to remove the single-lookback cliff risk that had killed other strategies, while staying on the non-prohibited equal-weight axis with only two tunables. The refinement backfired: it dropped the basket from Sharpe 2.17 to a break-even-negative signal (profit factor 0.968, CAGR -2.39%, four losing years), with its only positive contribution being a 2026 outlier-carried spike. The analyst abandoned it at the backtest-review gate on its first iteration, judging the robustness premise empirically falsified — the consensus gate over-filters and multi-horizon averaging smooths the edge away. The recorded lesson was that a sub-1.0 profit factor means the signal refinement itself failed and no sizing fix rescues it, holding the family line that only PF >1.2 baskets are worth optimizing.

Replacing the promoted basket's single trend with a 4-horizon ≥3-of-4 consensus does NOT improve robustness — it dilutes the net-directional momentum edge to profit factor <1.0 and negative CAGR (losing 4 of 6 years), the same sub-1.0-PF signature as the abandoned long-horizon and inverse-vol baskets; the family line holds that PF >1.2 with multi-year consistency (dual-TF confluence 1.211, cross-sectional 1.16) is a real edge worth iterating, while PF ≤1.0 means the signal refinement itself failed.

The analyst issued an 'abandon' verdict at the pre-optimization backtest-review gate, so optimization, analysis, and risk review never ran: the hypothesis's central claim — that a ≥3-of-4 fixed-horizon consensus improves robustness over the promoted single-trend basket — was empirically falsified, as the consensus gate and multi-horizon averaging smoothed the edge away to a sub-1.0 profit factor and negative CAGR that no parameter or sizing fix can rescue.

A long-short, equal-weight, net-directional momentum basket across ~10 Binance USD-M majors whose distinct mechanism was a MULTI-HORIZON TREND CONSENSUS signal — each leg goes long/short only when at least 3 of 4 fixed SMA horizons (20/60/120/240 days) agree, otherwise flat — rebalanced weekly via full-flatten-and-reopen, intended to extend the promoted single-trend Absolute-Momentum LS basket (Sharpe 2.17) by removing its single-lookback overfit/cliff risk while keeping the validated equal-weight scheme and only 2 tunables.

The consensus refinement degraded rather than improved the promoted basket: profit factor 0.968 (<1.0), CAGR -2.39%, Sharpe 0.12 with CI straddling zero (-0.72 to +0.71), expectancy -$19.6/trade, negative in 4 of 6 years (2022 -14.8%, 2023 -28.3%, 2025 -10.2%), 52.9% max drawdown — and the +16.9% headline was essentially all 2026 (+62%), carried by a few outlier prints (+44%, +38%, +32% on the final bar; kurtosis 168.9); per-leg sizing was sane (avg position 13.0%, no liquidation).