BinanceUsdmLongHorizonTSMomentumBasketMonthlyLS

Binance USD-M Long-Horizon Time-Series Momentum Basket (Established Majors, 6-Month Trend, Monthly Rebalance, Vol-Scaled, Long-Short)

A long-short, multi-instrument LONG-HORIZON (macro) time-series momentum basket across ~10 established, confirmed-full-history Binance USD-M majors (BTC, ETH, SOL, BNB, XRP, DOGE, ADA, LTC, LINK, BCH). It holds each major LONG when its OWN long-horizon trend (≈6-month) is up and SHORT when it is down, rebalancing only MONTHLY. This is the deliberate response to the session's hardest-won lesson: the analyst conclusively established that SHORT-HORIZON single-bar signals on majors are noise+fee dominated (both the 15m breakout-continuation and the 4H volume-climax-fade lost ~98%), and that the ONLY validated form of this primitive is DAILY-or-slower TREND (the daily vol-targeted basket: +392%, Sharpe 2.59). This strategy pushes that validated edge to the horizon where trend persistence is STRONGEST and noise/fees are WEAKEST — the classic 6-12 month managed-futures momentum — capturing the slow secular crypto trends with near-zero turnover. It is distinct from the weekly-rebalance promoted basket by horizon (months vs weeks → slower, lower-turnover sleeve) and fills the single most under-represented horizon bucket (macro_1w_plus at 0.2%). It uses only established-majors daily OHLCV — data-safe (broad-alt baskets die on data_unavailable) — and avoids every refuted class (no intraday, no MR fade, no basis/funding/options/COIN-M/HL/liquidations). Long-short improves the 87%-long skew; monthly rebalance makes fees trivial; few parameters resist overfitting.

Iteration-2 fix for the liquidation (cagr -100%, -232% day, ~11.5x gross, ~130% portfolio vol). Root cause was uncapped aggregate leverage: each of 10 highly-correlated majors independently targeted 30% annual vol with weight_cap 2.0 and no portfolio-level normalization, so the notional=equity*(1/n)*w formula summed to ~11.5x gross. Smallest-change fixes, keeping all passing signal logic and clean imports intact: (1) _compute_targets is now two-pass — Pass 1 computes the same signed vol-scaled equal-risk weights, Pass 2 rescales the entire book so Sum|notional| <= gross_cap*equity (1.0x), directly bounding aggregate leverage to ~1x gross and lowering portfolio vol into the ~15-25% target band; (2) target_vol lowered 0.30->0.15 and weight_cap 2.0->1.25 to further reduce per-leg risk before the cap; (3) a portfolio-level drawdown/liquidation guard tracks peak equity and flattens+halts if equity drops 30% below peak, so a sharp reversal cannot zero the account. leverage stays 1.0 (gross is capped at 1x in sizing, so no leverage_set_but_unused issue). This re-sizes the validated long-horizon TS-momentum primitive to sane vol for a clean re-evaluation of whether the 6-month edge is positive across the 2024/2025 regimes.

failed deflated Sharpe: DSR=0.0223, expected-max=1.0157 over 225 trials — selected Sharpe (0.496) indistinguishable from best-of-N noise, and BELOW the luck bar. Failing every robustness gate on final optimization attempt 2 of 2: is_significant=false, sharpe_ci_low=-0.1349 (CI straddles 0), PBO=0.545 (>0.5); probabilistic_sharpe 0.9722 vs DSR 0.0223 is the textbook PSR-vs-DSR selection-inflation trap. The walk-forward is is_overfitted=TRUE (avg IS 1.034 -> avg OOS 0.036 with OOS windows [1.57, 1.07, -2.54], the recent window deeply negative). The holdout 'pass' is SPURIOUS — its ratio 4.82 is inflated only because the WF-OOS baseline is near-zero (0.036), and holdout_sharpe 0.176 is negligible. The base edge is regime-decaying and outlier-carried: annual 2021 +60% collapsing to 2025 -15.5%, rolling Sharpe sliding from ~7 (2024) to -2.8 (2026-03), return_kurtosis 549 / skew 19. Not iterate (final attempt; the sensitivity surface is clean 0-cliff but the ENTIRE grid sits ~0.2-0.45, uniformly below the 1.0157 luck bar, so a further best-of-225 sweep re-selects a sub-significant, decaying config with no robust region to tune toward). Not revise_hypothesis: per the overfit->abandon rule this is high-IS/negative-OOS overfitting on a valid instrument set, and the validated daily/weekly vol-targeted momentum basket is already promoted — pushing the same primitive to a monthly long-horizon rebalance added no distinct promotable edge. FAILURE PATTERN: extending the promoted daily/weekly vol-targeted TS-momentum basket to a 6-month-trend / monthly-rebalance long-short does NOT inherit promotability — the monthly variant produces the same multiple-testing/overfit signature (DSR 0.022, optimized Sharpe below the luck bar, PBO 0.545, is_overfitted with a -2.54 OOS window) with a 2021-outlier-carried, post-2024 decaying edge; a clean 0-cliff sensitivity grid and a holdout ratio inflated by a near-zero baseline measure consistency, not significance or forward-generalization.

Long-short 6-month time-series momentum basket across 10 established Binance USD-M majors (BTC/ETH/SOL/BNB/XRP/DOGE/ADA/LTC/LINK/BCH) on daily bars, rebalanced monthly. Each leg is held long when its own ~180-day trend is up and short when down, vol-scaled to equal risk, then the WHOLE book is normalized to a portfolio-level gross budget (Sum|notional| <= 1.0x equity) so correlated legs cannot compound into runaway leverage. A hard portfolio drawdown guard flattens and halts the book if equity falls 30% below peak.

Implementation is sound: no liquidation, drawdown guard and portfolio gross-cap hold, turnover ~monthly (1.17), fees trivial ($559).

Vol targeting is calibrated — achieved annualized vol 15.6% ≈ 15% target; exposure_pct 1150% is a multi-leg reporting artifact (avg_position_pct 2.4%, effective book ~0.2x).

On the only validated primitive (daily-or-slower trend on majors) and fills the under-represented macro_1w_plus horizon bucket; PF 1.76, Sortino 1.08, Calmar 2.38 over the full sample.

benchmark_meaningful=false correctly handles the long-short basket; probabilistic_sharpe 0.957.

Return is 2021-concentrated (2021 +60%) with sustained RECENT decay: 2024 -0.4%, 2025 -15.5%, rolling Sharpe negative from mid-2025 to -2.8 by 2026-03.

Heavily outlier-driven (kurtosis 549, skew 19); a few days carry the result.

Sharpe 0.45 with sharpe_ci_low -0.28 straddling 0; information_ratio -0.63 — significance unproven.

Only 118 trades — thin for a walk-forward (though inherent to the monthly/6-month-hold macro horizon, not a code defect).

modest, CI straddles 0

near-zero as designed

sustained decay — holdout is the binding test

Clean sensitivity surface: 0 cliff parameters, gross-cap and drawdown-guard infra held (no blowup, not liquidated)

Genuinely market-neutral construction (benchmark_meaningful=false), low beta 0.16, sensible vol-scaling and monthly low-turnover design

Failed deflated Sharpe decisively: DSR=0.0223 (<0.95); optimized Sharpe 0.496 is BELOW the 225-trial expected-max luck bar of 1.0157 — selected config indistinguishable from best-of-N noise

is_significant=false, sharpe_ci_low=-0.1349 (CI straddles 0), PBO=0.545 (>0.5) — parameter selection more likely than not overfitting; PSR 0.9722 vs DSR 0.0223 is the textbook selection-inflation trap

Walk-forward is_overfitted=TRUE: avg IS 1.034 collapses to avg OOS 0.036 with one deeply negative OOS window (-2.54)

Holdout 'pass' (ratio 4.82) is spurious — inflated by a near-zero WF-OOS baseline (0.036); holdout_sharpe 0.176 is negligible

Severe regime decay: annual 2021 +60% -> 2025 -15.5%; rolling Sharpe from ~7 (2024) to -2.8 (2026-03); return_kurtosis 549, skew 19 — headline carried by single-day outliers

Fails promotion thresholds on every axis: Sharpe 0.45 vs 1.5, max_drawdown 40% vs 10%, win_rate 0.44 vs 0.45

Do NOT optimize — the strategy LIQUIDATED (cagr -100%, max_dd 87%, a -232% day on 2025-07-20, then all-zero returns). This is a fixable position-sizing/leverage design flaw, not necessarily a dead edge (the long-horizon TS momentum primitive is cited as validated on the daily sibling). Required fixes before any optimization: 1. ROOT CAUSE — uncapped portfolio leverage. The book runs ~11.5x gross (exposure_pct 1151%) and 130% annualized vol because each of the 10 assets INDEPENDENTLY targets target_vol=0.30 (30% annual vol) with weight_cap=2.0 and there is NO portfolio-level cap. Crypto majors are highly correlated, so 10 co-moving 30%-vol legs compound to ~130% portfolio vol. The notional = equity*leverage*(1/n)*w formula has no aggregate normalization. 2. FIX — target PORTFOLIO vol, not per-asset. After computing the per-asset vol-scaled signed targets, normalize the WHOLE book so the aggregate annualized portfolio vol is ~15-25% (account for cross-asset correlation, or as a simpler proxy cap total gross exposure Sigma|notional| <= ~1.0x equity). Drastically lower the effective leverage; reduce target_vol and/or weight_cap. Add a hard portfolio-level drawdown/liquidation guard so a sharp reversal cannot zero the account. 3. RE-BACKTEST AND RE-EVALUATE THE EDGE. Once sized to sane vol (~20% annualized, ~1x gross), check whether the 6-month TS momentum edge is actually positive across regimes — 2024 (-18.8%) and 2025 were negative even before the blowup, so the recent-regime edge is uncertain. Report Sharpe, max_dd, and annual returns on the sanely-sized book. DECISION RULE: sane sizing (max_dd < ~30%, no liquidation) + positive risk-adjusted edge across regimes (Sharpe > ~0.5) -> proceed to optimization. If, sized sanely, the long-horizon long-short momentum still loses or whipsaws in 2024/2025, the edge is not present on majors in recent regimes -> abandon (and note the promoted weekly cross-sectional/absolute momentum baskets already cover this primitive).

This strategy took the session's one validated primitive — daily-or-slower trend on majors — and pushed it to a 6-month, monthly-rebalanced, long-short vol-scaled basket, aiming to capture slow secular crypto trends at minimal fee and turnover cost. The implementation was clean (no liquidation, drawdown guard and gross-cap held, ~$559 fees, 0-cliff sensitivity surface), and the backtest passed the pre-optimization review with an 'optimize' verdict on a +89.5% return. But full optimization exposed it as best-of-N noise: a deflated Sharpe of 0.0223 with the optimized Sharpe 0.496 sitting below the 1.0157 luck bar, PBO 0.545, and a walk-forward whose in-sample 1.034 collapsed to 0.036 out-of-sample. On its second and final iteration the analyst abandoned it, concluding the edge was 2021-outlier-carried and post-2024 decaying, with no robust region to tune toward and no distinct promotable edge beyond the already-promoted daily/weekly basket.

Extending an already-promoted daily/weekly vol-targeted momentum basket to a 6-month-trend, monthly-rebalance long-short variant does not inherit promotability — it reproduces the same multiple-testing/overfit signature on an outlier-carried, regime-decaying edge.

The analyst abandoned it at the post-optimization ANALYZING stage (final iteration 2 of 2) because it failed every robustness gate: deflated Sharpe 0.0223 versus the 225-trial expected-max luck bar of 1.0157, PBO 0.545, is_significant=false, and a walk-forward flagged is_overfitted=TRUE (avg IS Sharpe 1.034 collapsing to OOS 0.036 with a -2.54 window), while the holdout 'pass' was spurious from a near-zero baseline.

A long-short, vol-scaled time-series momentum basket across ~10 established Binance USD-M majors that held each asset long or short on its own ~6-month trend and rebalanced only monthly, pushing the previously-validated daily-trend primitive to the strongest-persistence macro horizon with near-zero turnover.

The initial backtest returned +89.5% over 2019-2026 (2,114 data days) with Sharpe 0.45 (CI low -0.28), profit factor 1.76, 40% max drawdown and 118 trades; optimization produced Sharpe 0.496, +76.3% return, PF 1.86 and 36% max drawdown over 170 trades, but with heavy 2021 concentration and sustained recent decay (2024 -0.4%, 2025 -15.5%).