Cross-Exchange Funding Arbitrage

VERDICT: VARIANT RETIRED (g7 passed but other gates failed) + reconsideration note (3/11 gates pass)

Placebo (KILL SHOT): real PF 0.107 vs placebo p95 0.022 -> g7 PASS (frac of placebos >= real: 0.000)

Honest prior was 45% (highest of the batch, a real arb). Result below.

VENUE-DATA-AVAILABILITY

• Used (free + reachable): Hyperliquid, dYdX_v4
• UNAVAILABLE: Binance_global_fapi — HTTP 451 geoblocked
• UNAVAILABLE: Coinbase — spot only, no perp funding
• Coins: BTC, ETH
• Window: 2023-10-26 .. 2026-06-21 (HL+dYdX overlap; dYdX v4 funding starts 2023-10-31)
• DEGRADED: only 2 venues. A 2-venue spread is the minimum testable cross-venue arb. With only HL & dYdX, ‘most negative vs most positive’ reduces to long-cheaper / short-dearer of two legs. This is itself a finding: the free-data universe does not support a wide cross-venue book (Binance global geoblocked, Coinbase has no perp funding).

METHODOLOGY & DELTA-NEUTRAL CONSTRUCTION

• Per coin: long the venue with more-negative trailing 8h funding, short the more-positive. Matched notional => delta-neutral (same coin both venues; the two price legs cancel to first order, only the HL-vs-dYdX basis drift remains, which is second-order and not separately modeled — flagged as a small omitted term).
• Funding accrues HOURLY on both legs (both venues fund hourly). Side re-decided every 8h using the prior block’s realized spread (PIT, no look-ahead).
• Captured carry per hour = side * (f_HL - f_dYdX). PnL$ = capture * sleeve notional (equity/2 per coin, 1x, daily-compounded).
• Cost model (BASE, pessimistic): taker 4.5 bps + slippage 3.0 bps per side per venue. A side FLIP closes 2 legs + opens 2 legs = 30 bps of notional. BASE = hold-through (cost only on flip). CHURN variant pays full round-trip every reset.

HEADLINE METRICS (BASE hold-through)

• Final equity: $646 (start $5,000)
• Trades (8h blocks): 5812
• PF net: 0.107 Sharpe (daily, sqrt252): -16.47 MaxDD: -87.44%
• Months positive: 3% 2x-cost PF: 0.053
• CHURN variant: PF 0.000, final $1 (shows reset-churn cost drag)

11-GATE TABLE

Gate	Result	Detail
g1_min_100_trades	PASS	5812 trades (8h blocks)
g2_PF_ge_1.20	FAIL	PF 0.107 net
g3_Sharpe_ge_0.6	FAIL	daily Sharpe -16.47
g4_MaxDD_le_12pct	FAIL	MaxDD -87.44%
g5_pos_months_ge_60pct	FAIL	3% months positive
g6_bootstrap_LB_Sharpe_gt_0	FAIL	2.5% LB daily Sharpe -18.15
g7_placebo_beat_p95	PASS	real PF 0.107 vs placebo p95 0.022 (frac>=real 0.000)
g8_2x_cost_PF_gt_1.0	FAIL	2x-cost PF 0.053
g9_DSR_positive	FAIL	SR_hat -16.47 vs SR0 3.03, DSR 0.000
g10_no_component_gt_40pct	FAIL	max coin share 56%
g11_walkforward_OOS_ge_0.9IS	PASS	OOS PF 0.122 / IS PF 0.098 = 1.24

3/11 pass.

FUNDING-SPREAD COMPRESSION (early vs late)

• BTC: abs spread 0.195 -> 0.157 bps/hr (ann if captured 17.1% -> 13.7%); HL/dYdX funding corr 0.58
• ETH: abs spread 0.223 -> 0.162 bps/hr (ann if captured 19.5% -> 14.2%); HL/dYdX funding corr 0.46

COST SENSITIVITY

Scenario	fee/slip bps	PF	Final $	Sharpe
zero-cost	0/0	5.026	$6,347	12.52
optimistic	2.5/1.5	0.200	$1,878	-13.62
base	4.5/3.0	0.107	$646	-16.47
2x	9.0/6.0	0.053	$65	-18.24
3x	13.5/9.0	0.036	$7	-18.84

DEADBAND CADENCE SWEEP (POST-HOC — informs reconsideration, NOT gated)

Only reposition when |trailing 8h spread| exceeds a deadband; higher band = fewer flips = less cost but staler signal. This probes whether the REAL gross edge survives cost at any cadence.

deadband bps	gross %/yr	cost %/yr	NET %/yr	total flips
0	+8.99	85.97	-76.98	1520
1	+8.33	24.98	-16.66	441
2	+6.94	10.46	-3.51	184
5	+3.56	0.90	+2.66	15
10	+2.28	0.11	+2.16	1
20	+0.00	0.00	+0.00	0

Best survivable net ~ +2.66%/yr at 5bps deadband. Even fully churn-suppressed, the surviving edge is single-digit %/yr — far below the 5-7%/mo program target and below the PF>=1.20 gate.

PER-COIN ATTRIBUTION

• BTC: 44% of gross profit
• ETH: 56% of gross profit

WALK-FORWARD

• IS (‘2023-10-26’, ‘2025-05-30’) PF 0.098; OOS (‘2025-05-30’, ‘2026-06-21’) PF 0.122; ratio 1.24 (gate >=0.9)

BACKTEST-VS-LIVE DELTA

• Funding is realized historical (not assumed) — high fidelity.
• OMITTED/optimistic vs live: (a) HL-vs-dYdX basis drift on the delta-neutral legs not modeled (small but adds variance, slightly helps backtest); (b) dYdX v4 perp liquidity for BTC/ETH is thin vs HL — real slippage on a $2.5k sleeve leg is plausibly worse than 3 bps in stress; (c) funding-rate timing: we assume we hold across the funding stamp each hour — in live, missing a stamp by minutes flips a leg’s sign; (d) margin/liquidation buffer on each venue ties up capital not charged here; (e) spreads COMPRESS as arbs crowd (see compression section) — forward funding spread is likely <= the backtest’s.
• Net: live PF would be <= backtest PF. Treat backtest as an upper bound.

RECONSIDERATION NOTE (g7 passed => archetype not retired)

• The funding-spread SIGNAL is real: zero-cost PF ~5.0, Sharpe ~12, and the sign-permutation placebo is decisively beaten (0/200 placebos reach real PF). The cross-venue funding spread genuinely predicts captured carry direction.
• The archetype dies on COST, not on no-edge — the same mechanical failure as the single-venue HL funding harvest (whipsaw/churn drag). Gross ~+9%/yr is shredded by ~86%/yr of flip cost at signal-cadence; even an optimal deadband leaves only ~+2.7%/yr, sub-threshold.
• Reconsideration would require: (a) MORE venues (real >2-venue dispersion; free data blocked that here — Binance global geoblock, no Coinbase perp funding), (b) maker/rebate execution to cut the 30bps/flip, (c) a slow-cadence variant pre-registered fresh on OOS data (not the post-hoc deadband above). Absent those, this 2-venue HL/dYdX BASE config is RETIRED.