EthCrossVenuePositiveFundingCarrySpotPerpNeutral1H

ETH Cross-Venue Positive-Funding Carry — Long BINANCE_SPOT + Short USD-M Perp on Extreme Positive Funding (Delta-Neutral, Direct ETH Analog of the Promoted BTC Funding-Carry Survivor, 3-Parameter)

A DELTA-NEUTRAL, CROSS-VENUE funding-carry book on ETH that is a deliberate, near-exact ETH replica of the one cross-venue carry that has actually reached paper in this factory ('Cross-Venue BTC Funding Carry: Long Spot + Short Perp on Extreme Positive Funding', Sharpe 8.58). Mechanism: when the Binance USD-M ETHUSDT perpetual 8h funding rate runs EXTREMELY POSITIVE (longs paying shorts), open equal-notional LONG ETHUSDT.BINANCE_SPOT (cash leg, 1x, no leverage) + SHORT ETHUSDT.BINANCE perp (funding-collecting leg). While the position is held, the short perp COLLECTS the positive funding every 8h and the two legs are delta-neutral so directional ETH moves wash out. This directly addresses distilled-lesson L1 (cross-venue carry is the most reliable fee_edge death UNLESS per-cycle carry clears the ~30bps two-leg fee): entry is gated to funding >= +0.05%/8h (= 15bps/day), and historically ETH positive-funding episodes at that threshold persist 2-6 days, so the aggregate funding collected per round-trip has a median of roughly 40-60bps — comfortably above the ~0.30% round-trip cost (spot 0.20% + perp 0.10%). ETH is chosen over BTC because ETH perp funding hits extreme-positive far MORE frequently and with larger amplitude than BTC (more leveraged retail long flow, thinner dealer hedging), so the same proven mechanism runs on a richer, more frequent carry — giving a healthy trade count for robust walk-forward rather than the sparse-window death that kills low-frequency strategies. It fills the cross-venue quota gap (currently 6.9% vs 15% target) using a proven-survivor path, and keeps the parameter count to 3 (entry_funding, exit_funding, basis_stop) to fight the selection-inflation overfit that killed 202 prior hypotheses.

Fixes exactly the two linked findings from the semantic auditor, and nothing else. (1) Entry threshold: entry_funding is reverted to 0.0005 = 0.05%/8h in both the code default and config.parameters, matching the hypothesis text verbatim. Iteration 2 had raised it to 0.00075 on the basis of better per-cycle economics, but that changed which regime is under test (52 qualifying cycles at 0.0005 vs 28 at 0.00075) — a threshold that defines the edge is not mine to re-tune, so it is back to the specified value. (2) The contradiction between the hypothesis's economics and the code's own embedded measurement is resolved in favor of the measurement, and reframed: the docstring no longer presents the measured shortfall as justification for moving the gate, it presents it as evidence about the hypothesis. At the specified 0.0005 gate the measured median accrued carry is 0.203% of notional against a ~0.30% round trip — the median cycle is fee-negative by ~10bps and only 42% of cycles clear fees, roughly half the hypothesis's claimed 40-60bps, which flips the sign of the stated edge. The docstring also retains the regime evidence (ETH funding never exceeded 0.00010 in 2022, 2025 or 2026, so the entry regime has barely recurred since 2021) and notes the distribution is right-skewed, so mean-vs-median will diverge. Those are now surfaced for the analyst to rule on at BACKTEST_REVIEW, which is the correct decision point, rather than being papered over by a parameter change. Everything that already passed is untouched: imports unchanged (no banned modules), the continuous funding signal returned every bar in natural units, the equal-Decimal two-leg neutrality construction, the naked-leg guard, O(log n) cached funding lookups, the >= 8h staleness tolerance, spot-long-only compliance, per-cycle accrued-funding logging, and leverage = 1.0 with no leverage reference in sizing.

Delta-neutral cross-venue ETH funding carry. When Binance USD-M ETHUSDT 8h funding is >= +0.05%/8h (the hypothesis's gate), it opens equal-base-quantity SHORT ETHUSDT.BINANCE perp (collects funding) + LONG ETHUSDT.BINANCE_SPOT (neutralizes ETH delta), and closes both legs when funding normalizes to <= 0.01%/8h, on a basis-divergence stop, or on a 10-day holding cap. Both legs are sent with an identical Decimal quantity quantized to the coarser venue precision so net base delta is exactly zero by construction; every entry/exit logs both legs' quantity, price, net base delta and net USD delta, and each cycle logs its accrued funding as a % of notional against the ~0.30% two-leg round-trip cost. No leverage; stays flat if funding data is unavailable.

Abandon at BACKTEST_REVIEW rather than spending optimization budget on a gate that cannot fire in the OOS period. The hypothesis's two load-bearing claims — a 40-60bps median carry, and ETH funding hitting extreme-positive more often than BTC — are both contradicted by the collected data, so this belongs back with Research as a falsified premise, not a tuning problem.

ABANDON — unchanged from my prior review, and this iteration's own evidence strengthens it. I verified the ETH funding series directly last time against data/supplementary/fundingRate/ETHUSDT (7,029 prints, 2020-01-01 to 2026-05-31) and reproduced the docstring's year table exactly: 2022 max 0.00010 with 0.00% of prints >= 5bps, 2023 max 0.00071 with 0.46%, 2024 max 0.00102 with 2.28%, 2025 max 0.00010 with 0.00%, 2026 max 0.00010 with 0.00%. The restored gate of 0.0005 sits FIVE TIMES above the highest funding print recorded in 2022, 2025 or 2026 — the entry condition is not merely rare in those years, it is unreachable. Any walk-forward with recent OOS windows contains zero trades, and the final-15-day holdout (2026) contains zero trades: a hard promotion-gate reject. Reverting the threshold also moved the measured economics the wrong way exactly as predicted: avg_trade_return_pct fell from 0.029% at the 0.00075 gate to 0.0199% here, against the 0.15% floor. That matches the docstring's measurement that the median cycle accrues 0.203% against a ~0.30% round trip — fee-negative by ~10bps, only 42% of cycles clearing — roughly half the hypothesis's claimed '40-60bps median, comfortably above the ~0.30% round-trip cost', which flips the sign of the stated edge.

Weigh total_return, avg_trade_return_pct and end_unrealized_pct here, not Sharpe. Worth raising the daily-return construction for delta-neutral carry with the metrics owner as a systemic fix.

The reported Sharpe remains structurally misleading. This run shows sharpe_ratio 5.62 (CI 3.51-8.03, probabilistic_sharpe 1.0) alongside a NEGATIVE total_return of -3.01% and end_unrealized_pct of -9.31%. Sharpe is computed over daily returns dominated by small steady funding accruals while the negative total comes from the open pair's unrealized mark, so it rewards carry smoothness and ignores the loss. Note also max_drawdown_duration_days of 1033.62 against data_days of 309 — internally inconsistent. This distortion will recur on every delta-neutral carry book, not just this one.

Check cache.orders_open for both legs before submitting on either path, and have _close_pair retry an unflattened leg next bar rather than relying on the naked-leg guard to catch it at an arbitrary later price.

Still no in-flight-order guard (carried across two reviews). Entry keys off cache.positions_open() and exits submit without checking cache.orders_open, so between submission and fill a fresh bar can duplicate the pair or the exits. Backtest fills at next-bar open so it didn't manifest (74 clean pairs, long == short == 74), but live/paper latency is longer. The exit path does now log and warn on a residual leg rather than silently deferring to the naked-leg guard — an improvement — but still doesn't retry the unflattened leg.

The iteration-1 neutrality defect is genuinely FIXED. I verified leg-level trades: 2020-01-20 perp SHORT -$626.93 vs spot LONG +$593.68 at identical qty 181.236; 2020-01-28 perp SHORT -$3,102.42 vs spot LONG +$3,409.67 at identical qty 173.7999. The pair nets to roughly the funding accrual. beta -0.0003, benchmark_correlation -0.136 confirm the book is delta-neutral.

The semantic auditor's threshold-mismatch flag is also fixed — entry_funding reverted to the hypothesis's specified 0.0005, with the unfavorable measured economics reported as evidence rather than hidden behind a re-tune. That is the correct engineering call.

Funding IS being credited: $9,038 of funding collected across the run, accounted separately per cycle. The carry mechanism itself works as designed.

The mechanism has a proven sibling — the BTC cross-venue funding carry reached paper at Sharpe 8.58 — so this is not a speculative premise, it is a working mechanism aimed at the wrong target.

Edge does not clear costs: $9,038 gross funding vs $4,010 commission (44% of gross carry) plus $1,464 modeled impact (impact_cost_pct 25.8%). Net ~$2,737 on $100k over 6.5 years. avg_trade_return_pct 0.0261% vs the 0.15% floor; profit_factor 1.035 — textbook fee_edge signature.

The strategy's own measurement refutes the hypothesis: 52 cycles at the specified gate with MEDIAN accrued carry 0.203% against a ~0.30% two-leg round trip. The hypothesis claimed 40-60bps 'comfortably above' cost — the real figure is about half and flips the sign. Only 42% of cycles clear fees.

The entry regime is extinct. Entries by year: 2020=46, 2021=60, 2022=0, 2023=6, 2024=10, 2025=0, 2026=0. ETH 8h funding never exceeded 0.00010 in 2022, 2025 or 2026. Walk-forward OOS windows on recent data and the 15-day holdout would both contain ZERO trades — the 3-phase optimizer is structurally unable to evaluate this and would grade 2020-2021 as if it were robust.

Sharpe 2.08 rests on 60 active days inside a 2,392-day span, with sharpe_ci_low 0.2119 (CI nearly touches zero). Rolling Sharpe decays 9.06 (2021) to -1.94 (2024). max_drawdown_duration_days 1,276.

capacity_usd $1.50M — the edge exists only at toy scale, a promotion blocker independent of everything above.

beta ~ 0

Code↔hypothesis misalignment found by the semantic auditor — the code does NOT implement the hypothesis. Re-code the strategy to implement the hypothesis EXACTLY (instrument, timeframe, direction, the named edge/mechanic, sizing). Concrete issues: Entry threshold mismatch: the hypothesis text explicitly gates entry at funding >= +0.05%/8h ('entry is gated to funding >= +0.05%/8h (= 15bps/day)'), but both config.parameters.entry_funding and the code's own default (`float(p.get("entry_funding", 0.00075))`) are 0.00075 = 0.075%/8h — 50% higher. This is not cosmetic: the strategy's own measurement shows the two thresholds select materially different populations (52 cycles at 0.0005 vs 28 cycles at 0.00075), so the implemented strategy does not test the entry regime the hypothesis describes. The description acknowledges the change ('entry_funding is raised 0.0005 -> 0.00075') but the hypothesis text was never updated. | Hypothesis edge economics contradicted by the code's own embedded measurement at the same threshold: the hypothesis claims that at 0.05%/8h 'the aggregate funding collected per round-trip has a median of roughly 40-60bps — comfortably above the ~0.30% round-trip cost', while the module docstring and rationale report that at entry=0.0005 there are 52 cycles with MEDIAN accrued carry of 0.203% (~20bps) against the ~0.30% two-leg round trip, i.e. the median cycle is fee-negative by ~10bps and only 42% of cycles clear fees. The code's measured carry is roughly half the hypothesis's stated figure and flips the sign of the edge.

EthCrossVenuePositiveFundingCarrySpotPerpNeutral1H replicated the factory's one promoted cross-venue carry (BTC, Sharpe 8.58) onto ETH on the thesis that ETH's more frequent, larger positive-funding spikes would yield a richer carry. The engineering was honest and clean — perfectly delta-neutral by construction, funding genuinely credited ($9,038 collected), and, notably, the developer reverted a prior threshold re-tune and surfaced its own measured evidence that the carry falls short rather than hiding it. But that evidence was decisive against the hypothesis: median cycle carry of 0.203% fails to clear the ~0.30% two-leg round trip, avg per-trade return of 0.0261% sits far below the 0.15% floor, and the qualifying funding regime has essentially not recurred since 2021 (zero entries in the last 24 months). The reviewer ruled revise_hypothesis — a working mechanism pointed at a dead target — and the strategy was abandoned before optimization.

Faithfully porting a proven mechanism to a new underlying is worthless if that underlying's signal regime has gone extinct — ETH 8h funding hasn't exceeded the 5bps/8h gate since 2021, so the carry no longer clears the ~30bps cross-venue round trip and the median cycle is fee-negative, a structural fact no parameter choice can fix.

The backtest reviewer returned 'revise_hypothesis' at the pre-optimization gate: the mechanism was sound and correctly implemented, but aimed at the wrong target — ETH funding no longer reaches the required extreme, so recent walk-forward OOS windows and the holdout would contain zero trades, and the median cycle carry fails to clear the two-leg fee. The strategy's final stage is abandoned.

A delta-neutral, cross-venue funding-carry book on ETH — a near-exact replica of the promoted BTC funding-carry survivor — that shorted the Binance USD-M ETHUSDT perp against an equal-notional long in ETHUSDT.BINANCE_SPOT whenever 8h perp funding ran extremely positive (>= +0.05%/8h), collecting the funding while directional ETH risk washed out.

The base backtest confirmed the mechanism worked — genuinely delta-neutral (beta -0.0003), $9,038 of funding collected across 122 trades, Sharpe 2.08 and max drawdown 2.42% — but the edge did not clear costs: commissions were $4,010 (44% of gross carry) plus $1,464 modeled impact, netting ~$2,737 on $100k over 6.5 years, with avg_trade_return_pct 0.0261% against the 0.15% floor, profit factor 1.035, and toy capacity of $1.5M. The strategy's own measurement showed 52 cycles with median accrued carry of 0.203% versus a ~0.30% round trip (only 42% clearing fees), and the entry regime is effectively extinct (0 entries in 2022, 2025, 2026). No optimization stage ran.