PersistentDeepNegativeFundingLongCarryBasket8H

Persistent-Deep-Negative-Funding Long Carry Basket, Single-Leg Long-Only (BINANCE USD-M, 8 Liquid Perps, Go Long a Name ONLY When Its 8h Funding Has Been Deeply Negative for ≥3 CONSECUTIVE Stamps — Sustained Short-Crowding = Carry + Squeeze-Bounce — SINGLE-LEG So One Fee Not Two, Funding-Feed Dependent, Fee-Arithmetic Stated, 2-Parameter)

A LONG-ONLY, single-leg, funding-driven carry basket that respects the three lessons that just narrowed the viable design space to almost nothing: L30 (pure-OHLCV trend/momentum/reversion = 0/213, so a genuine NON-PRICE data dependency is required — this uses the FUNDING feed, the only deep/backfilled non-price series, NOT the shallow OI/taker/liquidation metrics feeds that data-walled prior attempts), L29 (no options), and L28 (delta-neutral TWO-leg carry fee-dies — so this is SINGLE-LEG, paying one round trip, and states the fee arithmetic). It is the persistence-gated evolution of the deployed negative-funding carry basket (Sharpe 1.97): instead of entering on an INSTANTANEOUS negative funding print, it requires funding to have been deeply negative for ≥3 CONSECUTIVE 8h stamps (24h of sustained short-crowding) before going long a name. That persistence filter is the distinct mechanism — it removes single-print noise entries, confirms a genuine crowded-short/squeeze setup, and guarantees the position is opened into SUSTAINED carry rather than a fleeting blip. FEE ARITHMETIC (per L28): entry requires funding ≤ −0.03%/8h persistent = −0.09%/day of carry COLLECTED while long (shorts pay longs); held ≥3 days that is ≥27bps of carry vs a single-leg USD-M round trip of ~0.09% (9bps) — a ~3× margin BEFORE the contrarian price bounce that typically follows crowded-short capitulation. Universe is 8 deep-history liquid USD-M perps (BTC, ETH, SOL, XRP, DOGE, LINK, AVAX, LTC), each gated INDEPENDENTLY, equal-weight, conservatively sized. NOT a two-leg spot-perp carry (single-leg), NOT a funding RANK rotation (independent per-name level+persistence gate, no cross-sectional sort → not the dead top-K funding rotation), NOT pure-OHLCV, NOT options. 2 core parameters (funding magnitude threshold, consecutive-stamp count).

Implements the hypothesis as the persistence-gated evolution of the deployed negative-funding carry basket. (1) Non-price data dependency (L30): the ONLY signal is the funding feed -- the deep, backfilled non-price series -- never OHLCV. (2) Single-leg (L28): one perp long per name pays one ~9bps USD-M round trip, and the stated fee arithmetic (>=27bps carry over a >=3-day hold at -0.03%/8h vs 9bps fee, ~3x margin, plus the bounce) is what clears the 0.15% avg-trade floor; the 24h min-hold prevents whipsaw before carry accrues. (3) No options (L29). (4) Independent per-name level+persistence gate, NOT a top-K funding rank rotation (no cross-sectional sort), so it is not the dead rotation class. (5) The 3-consecutive-stamp filter is the distinct mechanism vs the instantaneous-print deployed sibling -- it removes single-print noise and confirms genuine sustained short-crowding. Verification robustness: I confirmed against the real funding parquet that SOL (the primary, whose real funding_rates is the only per-leg series available in the Layer-3 sandbox) reaches <= -0.03%/8h for 3 consecutive stamps ~163 times in recent history (4.9% of stamps below threshold), so the sandbox reliably produces trades rather than looping on 'no trades'. In the full backtest, funding_rates_by_instrument supplies each of the 8 legs its own real funding series so all names trade. Long-only + leverage 1.0 + 10%/name sizing structurally avoids the 50%+ drawdown / -100% liquidation blowups that killed prior baskets (L15); the per-name price stop backstops the directional risk the single-leg (unhedged) structure carries. Venue is BINANCE USD-M futures because the edge is funding-based and requires holding the perp to receive funding.

Weak, decaying, and self-defeating despite good per-trade metrics. The aggregate is Sharpe 0.19 (CI [-0.49, 0.86] straddling zero) with total_return just 12.9% over ~5.8 years — barely productive — and the carry edge has decayed to ~zero since 2023 (2024 -0.07%, 2025 +1.3%, 2026 -2.2%) as funding normalized and deep-negative short-crowding setups became rare (only a handful of the 85 trades occur post-2023). Most decisively, the hypothesis's distinguishing innovation — the ≥3-consecutive-stamp persistence gate — HURTS the proven mechanism: it fires only 85 times in 6 years for Sharpe 0.19, versus the already-deployed instantaneous-negative-funding carry basket at Sharpe 1.97. The only optimization levers cannot fix this: reducing consecutive_stamps toward 1 simply converges to the deployed duplicate, and loosening entry_threshold below -0.03%/8h collapses the fee-arithmetic margin (27bps carry vs 9bps round-trip) into fragility. With a 0.19 base Sharpe and the edge concentrated in the faded 2021-2023 high-funding regime, the 3-phase optimizer would overfit that regime and fail the OOS/deflated-Sharpe gates. The funding-carry edge is already captured by the deployed sibling; this persistence-gated variant adds a filter that degrades it rather than improving it, so there is no promotable, non-duplicate configuration to tune toward. The excellent 9.9% drawdown and 1.88 profit factor do not compensate for a strategy that is barely productive and decaying. Not worth 2 hours of optimization.

Long-only, single-leg funding carry basket over 8 liquid BINANCE USD-M perps (SOL primary + BTC, ETH, XRP, DOGE, LINK, AVAX, LTC), each gated INDEPENDENTLY (no cross-sectional rank). Goes LONG a name's perp only after its 8h funding has been <= -0.03%/8h for >= 3 CONSECUTIVE stamps (24h of sustained short-crowding), collecting funding paid by crowded shorts and riding the squeeze-bounce that typically follows capitulation. Single-leg = one round-trip fee (not two). Exits when funding normalizes (>= exit_threshold, after a 24h min-hold), on a 15-day max-hold timeout, or a 12% protective price stop. No leverage, conservative 10%-equity-per-name sizing (<=0.80 max gross) so the book cannot liquidate and drawdown is bounded by spot moves. Funding is read per-leg from funding_rates_by_instrument (real per-leg series in the full backtest), with the primary leg falling back to the flat funding_rates dict (its real series in the sandbox); a leg with no funding data stays flat -- no price proxy is ever substituted.

At backtest-review, confirm total_trades and metrics_reliable on the FULL 8-leg run before trusting Sharpe/PF; if trades remain well below ~100, judge on carry economics + the deployed sibling's behavior rather than on noisy sample statistics.

Measurability risk. The sandbox produced only 2 trades over 334 days because Layer-3 exposes ONLY the primary (SOL) funding series; the other 7 legs stay flat there. In the full backtest funding_rates_by_instrument feeds all 8 legs, so trade count should be ~8x, but the entry is intrinsically low-frequency: a >=3-consecutive-stamp deep-negative gate fires in a few clustered episodes per name, and 24h min-hold / up-to-15d max-hold keep each name in-position across most qualifying stamps. Estimated full-backtest trades are roughly 60-100 across the basket over multi-year history -- borderline for statistical reliability. Not a code defect and there is no code fix (reducing the gate would violate the hypothesis's persistence mechanism), but the analyst should treat a sub-~100-trade full backtest as low-confidence rather than a validated edge.

No blocking change required. Optionally raise min_hold_bars toward the stated 3-day (9-bar) floor so the minimum hold clears the round trip on carry alone, ensuring even the shortest trades are not net-negative before the bounce.

Fee-arithmetic vs code drift. The hypothesis justifies fee viability on a '>=3-day hold accruing >=27bps of carry vs ~9bps round trip (~3x margin)'. The code's min_hold_bars=3 is 3 x 8H = 24h = ONE day, i.e. ~9bps of carry at the -0.03%/8h threshold -- roughly break-even on carry alone at the minimum hold, with the edge then depending on the extended (funding-normalization-driven) hold plus the squeeze bounce. The single-leg structure and per-stamp funding accrual keep this directionally sound (sandbox avg hold ~1d16h = ~15bps carry), but the stated '3x margin' only materializes on longer holds, not at the enforced floor.

Ignore sandbox PnL; require the full 8-leg backtest for any performance judgment.

Sandbox metrics are negative (total_return -0.57%, Sharpe -0.22, PF 0.68, avg_trade_return_pct -2.66%) but metrics_reliable=false with only 2 trades (one +$1032 win, one -$1519 loss). These do NOT show a code-defect signature (win_rate 0.50, |Sharpe| well under 5, PF not 0.0), so they are pure small-sample noise from the single-leg sandbox, not evidence of a polarity or sizing bug. Flagged only so the negative sandbox number is not mistaken for a validated result.

Genuine per-trade edge: profit_factor 1.88, avg_trade_return_pct 1.71% (clears fees), with excellent risk control (max_drawdown 9.9%, no liquidation)

Uses the deep/backfilled funding feed (funding_events_available 1039) — L30-compliant non-price dependency, no data wall; carry is credited so returns are real

Positive every year 2020-2023; proven mechanism family (deployed carry sibling Sharpe 1.97)

Aggregate Sharpe 0.19 with CI [-0.49, 0.86] straddling zero; total_return only 12.9% over ~5.8 years (~2%/yr) — barely productive

The carry edge decayed to ~zero post-2023 (2024 -0.07%, 2025 +1.3%, 2026 -2.2%) as funding normalized — deep-negative setups now rare

The persistence filter (the hypothesis's distinguishing innovation) HURTS: 85 trades / Sharpe 0.19 vs the deployed instantaneous-funding carry basket's Sharpe 1.97; loosening consecutive_stamps just converges to that deployed duplicate

Neither core param can fix it: loosening the funding threshold shrinks the fee margin into fragility; the edge is regime-decayed, so optimizing would overfit the 2021-2023 high-funding window

This strategy carefully respected the session's hard lessons — using the deep, backfilled funding feed as a genuine non-price dependency, staying single-leg to pay one fee not two, and stating its carry-vs-fee arithmetic — and it delivered real per-trade economics with a 1.88 profit factor, 1.71% per trade, and an excellent 9.9% drawdown. But its distinguishing innovation backfired: the ≥3-consecutive-stamp persistence gate fired only 85 times in six years for an aggregate Sharpe of 0.19, well below the deployed instantaneous-funding carry sibling's Sharpe 1.97, and the deep-negative-funding setups it needs have become rare as funding normalized post-2023. The analyst abandoned it at review — the funding-carry edge is already captured by the deployed basket, this persistence filter degrades it, and the two tunable parameters either recreate the existing strategy or collapse the fee margin, leaving nothing promotable to optimize toward.

Adding a persistence filter to an already-deployed funding-carry mechanism degraded rather than improved it — loosening the stamp count just converges to the existing sibling and loosening the threshold collapses the fee margin — and the deep-negative-funding edge itself has decayed since 2023, so a variant that filters a captured edge into near-zero productivity has no room to add value.

It was abandoned at the pre-optimization backtest-review gate: the aggregate Sharpe of 0.19 was near-zero, the carry edge had regime-decayed as funding normalized, and — decisively — the hypothesis's distinguishing ≥3-stamp persistence gate HURT the proven mechanism (85 trades / Sharpe 0.19 vs the already-deployed instantaneous-funding carry basket's Sharpe 1.97), so neither core parameter offered a promotable, non-duplicate configuration to tune toward.

A long-only, single-leg funding-carry basket over 8 liquid BINANCE USD-M perps (BTC/ETH/SOL/XRP/DOGE/LINK/AVAX/LTC), each gated independently to go long only after its 8h funding has been deeply negative (≤ -0.03%/8h) for ≥3 consecutive stamps — a persistence filter meant to confirm sustained short-crowding for carry plus a squeeze-bounce, paying one round trip and using the deep funding feed as its non-price signal.

It had genuine per-trade economics — profit factor 1.88, avg_trade_return_pct 1.71% (clears fees), 9.9% max drawdown, no liquidation, positive every year 2020-2023 — but was barely productive in aggregate: Sharpe 0.19 (CI straddling zero), total return just 12.9% over ~5.8 years (~2%/yr) across 85 trades, with the edge decayed to ~zero post-2023 (2024 -0.07%, 2025 +1.3%, 2026 -2.2%).