Trend + Carry Overlay (G10)

VERDICT: RETIRE ARCHETYPE — NO v2

2 / 11 gates pass. The kill-shot placebo gate (g7) FAILS decisively: real PF 0.783 vs sign-permutation placebo p95 1.234; 98.5% of random weekly-sign placebos beat the real strategy. Per the program rule (g7 fails -> RETIRE, no v2), this archetype is retired. The 6-month trend gate does NOT rescue carry. It is a worse version of the same failure: it trades ~55% less of the same losing edge and concentrates the loss into a DEEPER drawdown (-58% vs carry-v2’s -44%).

Honest prior was 30%. Result is consistent with the program’s hard-won lesson — overlays (reactive OR proactive) do not rescue a dead carry edge. This is the 8th archetype / Nth carry-flavored variant to fail.

---

What was built (methodology)

Long-only carry construction. Reused the exact point-in-time policy-rate panel (rates.py) and data (carry_hardened/data/*_D1.parquet, 2011-01..2026-06, weekly rebalance). Rank the 8 G10 currencies by policy rate; the carry “longs” are the top-3 yielders. Each long-yielder is expressed as a directional position in its USD pair via CCY_TO_PAIR sign (e.g. AUD = long AUDUSD; CAD = short USDCAD = long CAD). The low-yielders are NEVER shorted (anti-blowup design constraint). USD, when itself a top-3 yielder, has no non-USD leg and is skipped — so the long book naturally shrinks in the 2023-24 high-USD regime.

6-month trend gate (the new hypothesis), no look-ahead. For each candidate long-carry leg, the 126-trading-day (~6mo) moving average of the pair’s close is computed and lagged one day (ma.shift(1), close.shift(1) — uses only data through t-1). The leg is taken ONLY IF the pair’s price confirms the same direction the long-carry leg needs (long-ccy +1 pairs require close>MA; long-ccy -1 pairs require close<MA, i.e. the foreign currency trending up vs USD). If the trend disagrees, the leg is flat — never flips to short the low-yielder.

Costs / financing. Identical conventions to the hardened carry v2 engine: 0.5x spread (cost_mult) + 1.0 pip slippage per side, and a 75 bp/leg financing (swap-markup) drag baked into the base case. Vol-targeted (10% ann, 60d) across confirmed legs, capped at MAX_GROSS_LEV=30, $5,000 start equity.

Placebo. Sign-permute each weekly net P&L (random +/- per week, N=200), rebuild the equity curve, take its PF. The honest null for a long-only timing book: could a random long/flat timer with the same weekly P&L magnitudes produce this PF? Real PF must beat the placebo p95.

---

Headline metrics (BASE: trend gate ON, 75bp drag)

Metric	Value
Final equity (from $5,000)	$2,624
CAGR	-4.22%
Profit factor (weekly)	0.783
Sharpe (weekly, sqrt52)	-0.44
Max drawdown	-58.0%
Weeks / years	780 / 14.9
Avg legs held	1.10
Carry component	$+323
Price (FX) component	$-2,532
Total costs	$167

The carry component is genuinely positive (+$323) — the rate differential earns money. But the price/FX component (-$-2532) overwhelms it. The trend filter’s entire job was to keep the book on the right side of price; it failed completely.

---

11-Gate table

Gate	Result	Detail
g1_min_100_weeks	PASS	780 weeks
g2_PF_ge_1.20	FAIL	PF 0.783
g3_Sharpe_ge_0.6	FAIL	Sharpe -0.44
g4_MaxDD_le_12pct	FAIL	MaxDD -58.0%
g5_pos_years_ge_60pct	FAIL	25% years positive (4/16)
g6_bootstrap_LB_Sharpe_gt_0	FAIL	95% LB Sharpe -0.91
g7_placebo_beat_95pct	FAIL	real PF 0.783 vs placebo p95 1.234
g8_2x_cost_PF_gt_1.0	FAIL	2x-cost PF 0.779
g9_DSR_deflated_positive	FAIL	SR_hat -0.44 vs SR0 0.05, DSR=0.027
g10_no_ccy_gt_40pct	FAIL	max ccy share 57%
g11_walkforward_OOS_ge_0.9xIS	PASS	PF IS 0.704 / OOS 0.920

Note on the 2 “passes”: g1 is trivial (enough weeks exist). g11 “passes” only because the strategy loses about equally in-sample and out-of-sample (PF IS 0.704 / OOS 0.920) — both well below 1.0. A robustly-losing strategy passes a relative walk-forward gate; this is NOT evidence of edge, it is evidence of consistent failure. Every gate that tests for actual positive edge (g2, g3, g6, g7, g8, g9) fails.

---

HEAD-TO-HEAD: does the trend filter rescue carry? NO.

	Retired carry v2 (long+short)	Long-only carry, NO gate	Long-only TREND+carry (BASE)
PF	0.84	0.845	0.783
Sharpe	-0.29	-0.36	-0.44
Max DD	-44%	-62%	-58%
Avg legs	(3 long / 3 short)	2.44	1.10
Final $	—	$2,410	$2,624

The trend gate cut average legs from 2.44 to 1.10 (~55% fewer trades) yet PF moved the WRONG way relative to the no-gate long-only control on a risk basis, and max drawdown got DEEPER, not shallower (-58% vs -62% no-gate, vs -44% original carry-v2). This is the textbook “overlay trades less of the same losing edge, into a luckier- LOOKING but actually unluckier subsample” failure. The filter did not move the needle toward edge — it concentrated the book and deepened the loss.

---

Per-currency attribution (every leg loses)

Currency leg	Net P/L ($)
CAD	-1,361
GBP	-381
AUD	-319
EUR	-159
NZD	-157
JPY	+0
CHF	+0

Every single currency leg is a net loser. CAD (via short USDCAD) is the worst at $-1,361 — max ccy share 57%, blowing the concentration gate. JPY and CHF never traded at all: their long-ccy legs require USDJPY/USDCHF to trend DOWN, which (in a multi-year USD-strength regime) almost never coincided with JPY/CHF being top-3 yielders. The “diversified carry basket” collapsed to an essentially single-name (AUD/CAD) directional bet that lost.

---

Trade-count-reduction analysis: edge or just fewer/luckier trades?

The gate reduced activity by ~55% (avg legs 2.44 -> 1.10). If the trend filter captured real edge, the surviving trades should show materially better PF / Sharpe / shallower DD. They do not: PF 0.845 -> 0.783, DD -62% -> -58% (worse). Fewer trades, same-or-worse quality = the filter is removing noise symmetrically, not selecting winners. The placebo confirms it: 98.5% of random weekly-sign sequences with this strategy’s own P&L magnitudes produce a HIGHER PF than the real signal ordering. The signal ordering is worse than random.

---

Cost sensitivity

Base PF 0.783; at 2x costs PF 0.779 (g8 FAIL). The failure is structural, not cost-driven — even at zero incremental cost the PF sits far below 1.0 because the price/FX component is deeply negative. Cutting costs to zero would not lift this above breakeven.

---

Backtest-vs-live delta (what would be worse live)

• Rate panel granularity: policy rates hand-coded at decision dates; live a funded broker’s actual swap rates (not pure policy diff) drive carry — typically worse than the 75bp markup modeled, especially on exotics.
• Trend-gate whipsaw at the weekly grid: live, the 6mo MA cross would be evaluated intraday and could whipsaw more; the daily-lagged weekly evaluation here is mildly optimistic.
• Single-name concentration: live, a book that is 57% one currency (CAD) has gap risk (BoC surprises, oil shocks) not captured by daily bars.
• Net: a live deployment would do no better, likely worse than this -58% DD, -4.2% CAGR backtest. Not deployable.

---

Methodology concerns / honesty notes

• The trend gate is genuinely point-in-time (MA and close both lagged 1 day); no look-ahead. Verified the MA uses min_periods=MA_DAYS so no partial-window early signals.
• g11 (walk-forward) is the one non-trivial “pass” and it is misleading by construction: a relative IS/OOS PF ratio passes for any consistently-losing strategy. Reported transparently rather than counted as evidence of edge.
• The long-only design did its anti-blowup job in one sense (no short-squeeze blowup), but a -58% DD on a long-only book shows “never short” is not by itself protective when the long legs are simply wrong.
• No data fabricated, no config selected post-hoc. BASE config (trend ON, 75bp drag) was pre-registered before results.

Conclusion

The proactive 6-month trend entry gate is not different from the reactive overlays the program already retired — it is the same null result wearing a different hat. Carry’s rate-differential edge is real and small (+$323), but it is dwarfed by FX price losses that NO trend filter on this universe/window removes. RETIRE. No v2. Carry remains, as the memory states, with no surviving deployable variant.