PartiallyHedgedMomentumLS

Partially-Beta-Hedged Cross-Sectional Momentum, Long-Short (BINANCE USD-M, Weekly Long Top-3 Momentum Majors + Short HALF-Notional BTC Hedge — Threads the Gap Between Long-Only Momentum (Real Edge but 63% Drawdown = Floor-Fail) and Fully Beta-Neutral Momentum (38% DD but No Residual Edge): a HALF Hedge Keeps ~Half the Momentum Return While Halving the Drawdown Under the 50% Floor, 3-Parameter)

A REAL, codeable strategy that directly targets the ONE binding constraint now killing the only LIVE edge (cross-sectional momentum, deployed Sharpe 2.6): the >50% drawdown floor. The evidence brackets the solution precisely — my Crowding-Filtered long-only momentum variant had real return but 63% DD (floor-fail, L19), while my fully Beta-Neutral momentum overlay cut DD to 38% (under the floor) but killed the residual edge (too little momentum alpha survives a FULL hedge). This proposes the untried MIDDLE GROUND: a PARTIAL beta hedge. Weekly, long the top-3 relative-strength majors (the proven signal), and short a BTC hedge sized to only ~HALF the long book's BTC-beta — retaining roughly half the raw long-momentum return (which is real and fee-clearing) while cutting the market-beta drawdown roughly in half, targeting a sub-50% DD that clears the floor WITH edge intact. This is genuinely distinct from both my pending long-only momentum variants (this is hedged) and the dead fully-neutral overlay (this is HALF-hedged, keeping directional momentum exposure). It reuses the proven momentum signal (no L30 new-indicator risk), addresses the exact death cause with a single new parameter (hedge ratio), and if it fails it definitively resolves whether momentum can be salvaged under the DD floor. Long-short, 4 instruments (3 long + BTC hedge), weekly, USD-M. 3 core parameters (momentum lookback, number held, hedge ratio).

Directly targets the one binding constraint on the only live edge (cross-sectional momentum, Sharpe 2.6): the >50% drawdown floor. The evidence brackets the fix — long-only momentum had real return but ~63% DD (floor-fail, L19); a FULL beta-neutral overlay cut DD to ~38% but killed the residual edge. This implements the untried MIDDLE GROUND, a PARTIAL hedge, adding exactly one new parameter (hedge_ratio) to the proven signal (no L30 new-indicator risk). Mechanism in _rebalance: rank the 7 alt majors by momentum, hold equal-weight top-3 long (positive-momentum only), estimate each long's BTC-beta via rolling OLS of daily returns (verified the estimator recovers true beta), then short BTC at hedge_ratio x the beta-weighted long notional. BTC is deliberately excluded from the long universe and used solely as the hedge, avoiding any long/short-same-name conflict. The BTC hedge is adjusted toward its target via a single signed DELTA order (_set_btc_hedge) — never a close+reopen and bounded so it never flips past flat — minimizing hedge churn/fees; an empty long book (all-negative momentum) drives the hedge to zero so the book goes flat. VENUE: USD-M futures is required (the strategy SHORTS BTC), matching the deployed momentum winner. SIZING/LEVERAGE: sized to fixed equity fractions with gross ~0.90 of equity (long 3x0.20=0.60 + hedge ~0.5x0.60=0.30), so leverage stays 1.0 and is genuinely unused — no leverage-unused gate, no liquidation-prone amplification. Purely price-driven (no supplementary or tick data) so it avoids cross-instrument-supp and tick-timeout pitfalls and is timeout-safe (bounded per-bar work: momentum + a 30-point beta regression per long, weekly). All trading (longs + BTC hedge) happens in the rebalance routine via _submit_entry_instrument/close_position; the default primary entry path is neutralized (should_enter None, should_exit False, position_size 0) so the base never independently trades BTC. All 8 instruments are actively-traded deep-history USD-M majors, so every leg has data (no data_unavailable). 3 core parameters shared across names (not per-name tuned) to resist overfit.

failed holdout and deflated Sharpe: the holdout collapsed to Sharpe -0.10 (ratio -0.205, well under the 0.7 bar) — the strategy LOST money out-of-sample — while PBO is 0.833 (83% probability of overfitting) and DSR=0.0165 over 225 trials means the selected Sharpe is statistically indistinguishable from noise. Both the failed holdout and PBO>0.5 are unwaivable hard gates, so promote is impossible. The walk-forward is_overfitted flag is technically false only because the IS/OOS ratio (1.48/0.51 ≈ 2.9) squeaks under the 3.0 threshold, but the harder gates all agree it is overfit, and the OOS windows [-0.45, 1.69, 0.28] are inconsistent. Even optimized, Sharpe is 0.645 (below the 1.5 floor) and the return is still concentrated in 2021 (+462%) with 2025 (-7.4%) and 2026 (-10.8%) negative. I flagged this exact risk at backtest review ('watch the PBO/deflated-Sharpe gates and abandon rather than promote if the 2024-2026 OOS collapses; every Sharpe-0.7 momentum candidate this session cleared review but failed these gates out-of-sample'), and it failed precisely that way. This is the overfit signature (failed holdout, high PBO, DSR ~0, single-regime concentration), which per policy is abandon not iterate — a second optimization pass would re-select the same 2021-concentrated overfit. The calibration is decisive: the entire cross-sectional momentum family this session is a graveyard (dollar-neutral spread -96% and rank-weighted -91% both fee_edge, equity-breaker and beta-neutral overlay overfit, risk-parity risk_reject), and this partial-hedge port — the best-behaved of them — still fails the holdout. Abandon.

Partially-beta-hedged cross-sectional momentum, long-short on BINANCE USD-M. Weekly it ranks 7 alt majors (ETH/SOL/BNB/XRP/ADA/AVAX/LINK, 1-DAY bars) by trailing relative strength and holds an equal-weight top-3 long (absolute-momentum gate: only positive-momentum names, else cash), then SHORTS BTC sized to hedge_ratio (default 0.5) of the long book's BTC-beta-weighted notional, where each long's beta is a rolling OLS of daily returns vs BTC. BTC is the dedicated hedge instrument (excluded from the long set). The half hedge keeps ~half the directional momentum return while roughly halving the market-beta drawdown, targeting sub-50% DD with edge intact. Pure-price, no supp data, no ticks. 3 core parameters: momentum_lookback, top_k, hedge_ratio.

Confirm the USD-M backtest venue uses NETTING OMS (default here). Under hedging OMS, replace BUY-to-reduce with an explicit position-reduce/close.

The BTC hedge is managed by a signed DELTA market order via _submit_entry_instrument (SELL to add short, BUY to reduce short). Correct ONLY under NETTING OMS, where a BUY against a net-short BTC position covers it. Under HEDGING OMS the BUY-to-reduce path would open a separate long and _position()'s positions_open[0] could return the wrong leg. Evidence it's fine: factory's standard positions_open[0] pattern assumes netting, and the sandbox executed 4 BTC shorts (short_win_rate 0.75) with DD cut to 24.9% as the partial hedge intends. Flagged for confirmation only.

Analyst: hedge resolves the DD-floor question (sub-50% achieved) but longs didn't clear positive return this window. BACKTEST_REVIEW abandon-or-optimize call; do not re-code.

PERFORMANCE FLAG FOR BACKTEST_REVIEW (not a QA defect). Sandbox negative: total_return -20.0%, PF 0.44, win_rate 0.375, Sharpe -0.80, avg_trade_return_pct -4.15%. But the hedge DID meet its goal — max_drawdown 24.9% (well under the 50% floor and below the long-only siblings), short_win_rate 0.75 vs long_win_rate 0.32. The partial-hedge mechanism works as specified; the residual negative return is the crypto-majors momentum edge failing this window, same as siblings.

If alignment isn't guaranteed, gate the snapshot on all legs sharing ts_event before computing momenta/betas.

Leg-alignment assumption: calculate_signal computes momentum + rolling-OLS beta assuming an alignment barrier buffered all 8 same-tf legs for that timestamp. If same-ts_event alignment isn't guaranteed, a leg could be one bar stale, slightly shifting ranks/betas. Would not break the hedge or flip polarity.

Validated its thesis: the ~half BTC-beta hedge cut max_drawdown to 38.1% (under the 50% floor, vs 63% long-only) while RETAINING edge (PF 1.374, avg_trade_return_pct 12.6%, +909%)

2022 only -25.3% (vs -55%/-58% for the unhedged momentum variants) — the hedge genuinely halved the market-beta drawdown

Functions correctly (211 trades, full return series); reuses the deployed proven momentum signal; hedge_ratio is a single well-motivated new tunable; Sharpe 0.70 with positive CI low (0.05)

Single-regime concentration: 2021 +659% (and 2024 +81%) carry the headline; 2022/2025/2026 negative, rolling Sharpe decayed from 8.5 to -0.56 — high deflated-Sharpe/PBO risk

Sharpe 0.70 below the 1.5 floor; CI low 0.053 weaker than the spot-momentum sibling's 0.20

Close sibling of the spot-momentum variant already sent to optimize (both ~Sharpe 0.7, DD 38%, 2021-concentrated) — running both is partly redundant

A weaker port of the deployed momentum winner (Sharpe 2.6)

Best-behaved momentum-family variant: the half-hedge held optimized max_drawdown to 30.4% (under 50%) with PF 1.35 and 367 trades

walk-forward is_overfitted flag technically false (IS/OOS ratio ~2.9, just under 3.0)

HARD FAIL: holdout failed — holdout_sharpe -0.10, ratio -0.205 (lost money out-of-sample)

HARD FAIL: PBO 0.833 > 0.5 — 83% probability of overfitting

deflated_sharpe 0.0165 << 0.95, is_significant=false, sharpe_ci_low 0.018 — selected Sharpe indistinguishable from noise; OOS windows [-0.45, 1.69, 0.28] inconsistent

Optimized Sharpe 0.645 below the 1.5 floor; return still 2021-concentrated (+462%) with 2025/2026 negative

Entire momentum family is a graveyard (dollar-neutral spread -96% fee_edge, rank-weighted -91% fee_edge, equity-breaker overfit, beta-neutral overlay overfit, risk-parity risk_reject)

PartiallyHedgedMomentumLS was the best-behaved of the session's many momentum-salvage attempts: by shorting a half-notional BTC hedge against the long top-3, it hit its explicit target — cutting max drawdown to 38.1% (30.4% optimized), under the 50% floor, while retaining edge (PF 1.374, +909%) — and earned an 'optimize' verdict. But optimization revealed the same fatal overfit as its siblings: the holdout lost money (Sharpe -0.10), PBO was 0.833, deflated Sharpe was 0.0165, and the return remained concentrated in 2021 with 2025/2026 negative. The analyst abandoned it on the unwaivable failed-holdout and PBO gates, noting the reviewer had flagged exactly this risk. As the hypothesis itself intended, its failure definitively resolved that the cross-sectional momentum family cannot be salvaged under the drawdown floor this session — even the variant that structurally beat the floor still could not survive out-of-sample.

The partial hedge genuinely solved the drawdown problem (30-38% DD with edge intact), proving the mechanism works, but it could not solve the deeper issue that the momentum edge is concentrated in the 2021 bull and does not generalize — a strong in-sample Sharpe that fails holdout with high PBO and DSR ~0 is overfit and should be abandoned, definitively resolving that momentum cannot be salvaged under the drawdown floor this session.

It passed backtest-review (verdict: optimize) but was abandoned at the post-optimization analysis stage on unwaivable hard gates: the holdout lost money out-of-sample (Sharpe -0.10, below the 0.7 bar), PBO was 0.833 (83% overfit probability), and DSR ~0.017 made the selected Sharpe indistinguishable from noise — the same overfit signature that killed every Sharpe-0.7 momentum candidate this session.

It reused the deployed cross-sectional momentum signal in a partially-hedged long-short form — weekly long the top-3 relative-strength alt majors plus a short BTC hedge sized to only ~half the long book's BTC-beta — seeking the middle ground between long-only momentum (real edge but 63% drawdown) and a full beta hedge (38% DD but no residual edge), keeping ~half the momentum return while halving the market-beta drawdown under the 50% floor.

It validated its drawdown thesis on the initial backtest: 211 trades, +909% return, profit factor 1.374, avg_trade_return_pct 12.6%, Sharpe 0.70, and max drawdown 38.1% (2022 only -25.3% vs -55%/-58% for unhedged siblings); but optimization exposed severe overfit — holdout Sharpe -0.10 (ratio -0.205), PBO 0.833, deflated Sharpe 0.0165, and returns still concentrated in 2021 (+462%) with 2025/2026 negative.