NegativeFundingCarryLongBasket5Perps4H

5-Asset Crypto Majors 4H Negative-Funding-Rate Long-Only Carry Basket (BTC+ETH+SOL+AVAX+ADA on BINANCE USD-M, NO Trend Filter, 3-Parameter, Basket-Diversification, Direct Implementation of Analyst Recommendation)

A LONG-ONLY multi-instrument carry-collection basket strategy on 5 BINANCE USD-M perpetuals: BTCUSDT, ETHUSDT, SOLUSDT, AVAXUSDT, ADAUSDT. Mechanism: at each 4H bar close, for each asset in the basket, check the most recent 8h funding rate (received from supplementary data feed). If funding rate is below funding_entry_threshold (default -0.02%/8h = -0.06%/day equivalent, meaning shorts are paying longs), OPEN LONG to capture the funding payment stream. Exit when funding regime turns positive (regime reset) OR after max 48-bar (~8 day) time stop. Per-asset position 15% (basket sizing across 5 assets = 75% max gross exposure). STRATEGIC RATIONALE: (1) DIRECT IMPLEMENTATION OF ANALYST RECOMMENDATION — the BTC Persistent Negative Funding Carry Long abandon notes explicitly recommend 'The right place to test this mechanism is 4H bars on a basket of crypto majors (BTC + ETH + SOL) where the conjunction probability is higher and individual asset noise is averaged' AND 'Future hypotheses in this class should drop the SMA filter entirely and use a different protection'. This proposal implements BOTH recommendations: 4H bars on a 5-major basket, NO trend filter, NO SMA gate — the failing conjunction (negative funding + above SMA, ~1 trade in 6 years on daily BTC) is structurally avoided. (2) MECHANISM PROVENANCE — funding rate carry is among the most mechanically-guaranteed cash flows in crypto: when funding < 0, shorts pay longs every 8h. Unlike directional strategies (Donchian, momentum) which depend on price prediction, funding carry has documented edge because it's a STRUCTURAL transfer between counterparties. The mechanism's edge is the funding payment itself, not a trade thesis about future price direction. (3) BASKET DIVERSIFICATION SOLVES THE TRADE-COUNT PROBLEM — single-asset BTC daily produced 1 trade in 6.4 years; with 5 assets at 4H and no trend filter, expected ~50-150 trades/year (negative funding occurs ~10-30% of time per asset, 5 assets × 4H bars × 365 days × probability ≈ 60-120 per year). Comfortably above the 5-trade walk-forward floor. (4) PORTFOLIO QUOTA FILL — multi-instrument scope at 8.9% (target ≥10%), this adds a 5-asset basket. Long-only at 87.5% over-represented but the basket structure is structurally different from single-asset long-only (5 partial positions vs 1 full position, basket-level Sharpe vs single-asset Sharpe). Funding-rate signal class has had zero successful basket implementations in the 1101-experiment portfolio — this opens an under-explored signal class. (5) AVOIDS ALL FAILED MECHANISM CLASSES: NOT Donchian-breakout (system-wide regime decay confirmed: BTC, AVAX, XRP, SOL all abandoned); NOT funding+SMA (conjunction failure confirmed 3+ times); NOT premium-index basis MR (cliff failure); NOT cross-venue basis arb (fee math failure); NOT skewness (mechanism class refuted); NOT OI surge (regime dependent); NOT volume surge daily (conjunction failure); NOT Connors down-day (deprecated); NOT Bollinger MR (cliff-prone); NOT BNB single-asset (moratorium, but BNB is intentionally NOT in this basket — replaced by ADA). (6) ITERATION-SAFE: only 3 hypothesis-declared tunable parameters (funding_entry_threshold, min_consecutive_funding_periods_negative, position_pct_per_asset). NO stop_loss parameter, NO take_profit parameter — explicit fix for LINK/ETH iter-2 optimizer-reversal failure. Risk is bounded by (a) funding regime reset exit (when funding turns positive, exit immediately), (b) hard 48-bar time exit (NOT tunable), (c) 15% per-asset sizing × 5 max = 75% max gross. The optimizer CANNOT push extreme parameter values because there are no risk-tunable parameters. (7) FEE MATH EXPLICITLY VIABLE: at funding -0.02%/8h, ~3 funding events captured per trade (avg 2-5 day hold) = ~0.06% funding collected per trade. Plus typical directional gain when funding regime ends with positive price move (correlated with negative funding squeeze): 0.5-1.5% per trade. Gross per winning trade: ~0.6-1.6%. Per losing trade (funding stays negative but price falls): bounded by 48-bar time exit + price drop typically 1-3% × 15% sizing = 0.15-0.45% per losing trade. Net of 0.10% RT fees: 0.5-1.5% per winner, -0.25 to -0.55% per loser. At 55-65% win rate: EV ≈ +0.5-0.9% per trade NET. ANNUAL: ~80 trades × 0.7% net = ~56% gross basket return at 15% per-asset sizing (across 5 assets = ~11% per-asset annualized). Sharpe target 1.0-1.8.

Iteration 3 fixes ONLY the reported Layer-3 failure (no trades) — Layers 1 and 2 already passed and the entry/exit mechanism, sizing, locked constants and signal logic are untouched. Diagnosis from the data, not guesswork: the Layer-3 sandbox injects supplementary data for the PRIMARY instrument only (it builds the flat `funding_rates` series; `funding_rates_by_instrument` is only assembled on the real backtest path), so in the sandbox the 5-asset basket collapses to its primary leg. I measured the funding parquet over the sandbox window (last 365 days): BTCUSDT NEVER reached the -0.02%/8h entry threshold (0 events, min -0.0152%), which is exactly why 0 entries fired, while SOLUSDT had 36 qualifying events spread over 22 distinct days (AVAX 70, ADA 20, ETH 4). The fix is therefore a one-line config change: SOLUSDT becomes the primary/reference leg and BTCUSDT moves into extra_instruments. Basket membership, threshold, sizing, exits and the mechanism are all unchanged — the real backtest still funds every leg via funding_rates_by_instrument — but the sandbox now has a primary leg whose funding actually crosses the threshold, so it produces trades. Second, contained change: `_latest_funding` previously rescanned every key of the funding dict on every bar for every one of the 5 legs (O(n) per bar, ~1.2B key conversions over the full 8-year multi-leg backtest) — a near-certain wall-clock timeout at the BACKTESTING stage even though the sandbox survived it. Each leg's series is now parsed once into a sorted (ts, rate) array and read with bisect_right (O(log n)), with identical causal 'last event at-or-before ts, within 12h' semantics. `bisect` is stdlib, so Layer 1 stays clean. No fallback edge was added: a leg with no funding series still simply does not trade.

failed deflated Sharpe: DSR=0.0001, expected-max=1.80 over 225 trials — the selected config is a best-of-N noise selection, confirmed by PBO=0.7103 (>0.5, hard fail) and a catastrophic holdout collapse to Sharpe 0.317 (ratio 0.084 vs required 0.70, hard fail). The funding-carry mechanism is genuinely real and credited (positive alpha 0.063, low beta, positive every year, avg_trade_return_pct >2%) — this is NOT a fee-edge or an uncreditable-funding death — but the honest full-sample Sharpe is ~0.53-0.59 with a CI that straddles zero, far below the 1.5 promotion floor; the WF-OOS windows (avg 3.79) were lucky draws inconsistent with both the full sample (0.59) and the final-window holdout (0.32). Both tunable parameters were fully swept with no robust region to tune toward, so iterate cannot lift a genuine ~0.5-Sharpe edge past the robustness gates, and the two hard failures make promote impossible. Not revise_hypothesis: the carry premise is sound and a sibling of this exact construction has passed elsewhere — the failure here is overfit parameter selection on a modest true edge, not a dead premise needing reframing.

Long-only negative-funding carry basket across 5 Binance USD-M perpetuals (SOL, BTC, ETH, AVAX, ADA) on 4H bars. For each asset independently: open a 15%-of-equity LONG when the most recent 8h funding rate is below -0.02% (shorts paying longs) and the funding period is negative; exit immediately when funding flips positive (regime reset) or after a hard 48-bar (~8 day) time cap. No trend/SMA filter and no tunable stop-loss/take-profit — only two tunable parameters (funding_entry_threshold, position_pct_per_asset). Max gross exposure ~0.75x equity, leverage 1.0.

Verification failed (Layer 3 — sandbox backtest): No trades produced Bar type used: BTCUSDT.BINANCE-4-HOUR-LAST-EXTERNAL, Bars processed: 2181 Diagnostics: should_enter() returned a side 0 times over 2177 evaluated bars -> your ENTRY CONDITION never triggered. Loosen the entry logic / thresholds. Ensure your strategy produces trades with the given data and parameters.

Analyst to judge whether the naked-long directional exposure has positive expectancy across regimes on the full 5-asset basket. The 'mechanically-guaranteed cash flow' framing does not hold for an unhedged long — the carry alone is fee-negative.

This is NOT delta-neutral funding carry — it is a NAKED LONG-ONLY directional bet gated on negative funding. There is no hedge leg, so the dominant PnL driver is each perp's price direction, not the funding stream. The hypothesis's own fee math concedes the pure carry is ~0.06%/trade, BELOW the ~0.10% round-trip fee, so profitability depends ENTIRELY on the directional thesis that negative funding precedes recoveries. Sandbox bears this out: break-even (avg_trade_return_pct -0.007%, PF 1.002) with inverted realized reward:risk (avg_win $475 ~+3.2% vs avg_loss $770 ~-5.1%). The moves clear fees (not a sub-fee-floor failure), but there is no directional EDGE in-sample. Whether 'negative funding -> bounce' is real is an analyst question over the full basket.

Analyst/backtest to confirm the full 5-asset basket produces ~100 trades/yr AND clears the floor. Do not read the 21-trade SOL-only sandbox as the basket's edge or trade count.

The sandbox is UNREPRESENTATIVE of the basket. Per the developer's own diagnosis, the Layer-3 sandbox only assembles the flat funding_rates series for the PRIMARY leg (SOL); funding_rates_by_instrument is empty in the sandbox, so only the SOL leg trades. The 21-trade / break-even result is a single-asset (SOL) window, not the 5-asset basket. Both trade count (dev estimates ~100/yr) and edge must be judged on the full backtest where every leg is funded.

Round qty to each instrument's size precision before submission, mirroring the single-asset strategies.

Per-asset order qty is submitted as Decimal(str(qty)) with NO rounding to each instrument's size precision (BTCUSDT 0.001, ETHUSDT 0.0001, etc.). An unrounded high-precision qty can be rejected or silently re-rounded, distorting sizing on the non-SOL legs. The sandbox survived (SOL-only) so this was not exercised on the other legs.

Positive in all 7 years including the 2022 crash — regime-robust; low beta (0.065) + positive alpha (0.063) show it's credited carry, not disguised long beta

profit_factor 1.84, avg_trade_return_pct 1.47% (~10x floor), expectancy +$375/trade, favorable reward:risk (avg_win > avg_loss), positive skew +1.57

383 trades across the basket — the 4H/no-SMA design solved the prior single-asset trade-count problem; ample sample for optimization

Funding is credited (2,751 events); differentiated market-neutral-ish carry, an under-explored signal class in the portfolio

Sharpe only 0.53 with CI [-0.093, 1.22] marginally straddling zero — positive but not yet significant; returns lean on 2021/2023

end_unrealized_pct -38% — open longs underwater at backtest end (no-stop design rides losers to the 48-bar cap); depresses headline rather than inflating it, but flags risk management

information_ratio -0.54 vs equal-weight basket (mitigated: only ~14% avg exposure, so Sharpe is the fair read)

Genuine credited funding-carry mechanism: low beta (0.065), positive alpha (0.063), positive in all 7 years including the 2022 crash — this is real carry, not disguised long-beta and not a fee-edge death (avg_trade_return_pct 1.47% raw / 2.12% optimized, ~10x the futures fee floor)

Optimized profit_factor 2.42, max_drawdown 7.1%, sensitivity grid clean (no cliffs, Sharpe stable across the 5x5 heatmap)

HARD FAIL — PBO 0.7103 > 0.5: the parameter selection is more likely than not overfit

HARD FAIL — holdout collapsed to Sharpe 0.317 (ratio 0.084 vs required 0.70); the honest final-window read is ~0.3, not the 3.79 WF-OOS average

Deflated Sharpe 0.0001 (needs 0.95) over 225 trials — selected Sharpe indistinguishable from best-of-N noise; is_significant=false; Sharpe CI [-0.05, 1.19] straddles zero

Full-sample headline Sharpe is only 0.53/0.59 — the WF-OOS windows [5.33, 2.65, 3.38] were lucky draws inconsistent with both the full sample and the holdout

information_ratio negative (-0.54/-0.61): underperforms the equal-weight basket benchmark on a risk-adjusted basis

Detailed feedback written to workspace/discussions/35bca890-127a-41fa-83cb-342dc24ce94f/iteration_1_analyzing_feedback.md. Summary: This is the iterate exception, not a clean DSR abandon — the walk-forward is genuinely not overfit (avg OOS 4.06 > IS 2.33, all OOS windows positive), sensitivity is cliff-free, and the holdout is POSITIVE in absolute terms (+1.79), failing only the relative ratio. The DSR 0.477 / PBO 0.8135 failure is substantially a multiple-testing artifact of an OVER-BROAD sweep: 2 of the 4 optimized parameters are completely inert (min_notional and min_consecutive_funding_periods_negative show identical Sharpe across all variations), so 225 trials taxed a ~2-effective-dimension search and pushed the optimized Sharpe (2.39) just under the expected-max luck bar (2.43). Specific changes for attempt 2: (1) DROP the two inert params from optimization — fix min_notional=5.0 and min_consecutive_funding_periods_negative=1 as constants; optimize ONLY funding_entry_threshold and position_pct_per_asset. (2) CUT the trial budget hard (~30-50 trials for 2 dims) so the expected-max luck bar reflects the real 2-D search. (3) RESTRICT the search to the stable plateau sensitivity already mapped: funding_entry_threshold in [-0.00024, -0.00018], position_pct_per_asset in [0.12, 0.18]. (4) Keep the no-stop/no-TP design (correct). HARD ABANDON-CRITERION FOR ATTEMPT 2 (do not iterate a third time): the real risk is regime decay, not parameter fragility — trades concentrate in 2021's negative-funding era and thin out in 2024-2026. Attempt 2 must clear BOTH deflated_sharpe >= 0.95 / is_significant=true AND a positive holdout keeping >= 70% of WF-OOS Sharpe. If it still fails DSR or the recent holdout still degrades below 0.70, the negative-funding carry has structurally decayed on this universe -> abandon (NOT revise_hypothesis; the decay is a market-wide funding-regime fact, not a wrong-target problem).

This basket directly implemented a prior analyst recommendation — test the funding-carry mechanism on 4H bars across a basket of majors with no SMA filter — and it genuinely fixed the earlier single-asset trade-count problem, producing 383 trades and a real, credited carry edge that stayed positive through every year including the 2022 crash. But its headline Sharpe was only ~0.53 with a confidence interval straddling zero, so it was sent to optimization on the strength of its mechanism rather than its risk-adjusted return. There the statistics turned decisive: PBO 0.71, a deflated Sharpe of 0.0001, and a holdout that collapsed from a flattering 3.79 walk-forward OOS average to just 0.32 exposed the selected parameters as best-of-N noise. With no robust region to tune toward and two unwaivable hard failures, the analyst abandoned it after three iterations — noting the carry premise itself is sound (a sibling construction passed elsewhere), just overfit on a modest true edge here.

A mechanically real edge (credited funding carry, positive every year) can still be unpromotable when its true Sharpe sits near 0.5 with no robust parameter region — stellar walk-forward OOS numbers that diverge sharply from the full sample and holdout are a luck/overfit signal, not a green light.

The analyst abandoned it at the post-optimization ANALYZING stage on two hard failures: PBO 0.7103 (>0.5) and a holdout collapse to Sharpe 0.317 (ratio 0.084 vs the required 0.70), with deflated Sharpe 0.0001 over 225 trials and an optimized Sharpe of just 0.59. The walk-forward OOS windows (avg 3.79) were lucky draws inconsistent with both the full-sample ~0.5 Sharpe and the honest final-window holdout of ~0.32.

A long-only, no-trend-filter negative-funding carry basket that went long any of five Binance USD-M majors (BTC, ETH, SOL, AVAX, ADA) on the 4H bar whenever its 8h funding rate fell below a negative threshold, collecting the funding stream and exiting on a funding-regime flip or a 48-bar time cap.

The default backtest was positive but weak on a risk-adjusted basis: total return ~105.5% over ~7 years with 383 trades, win rate 60.3%, profit factor 1.84, and avg_trade_return_pct 1.47% (well above the fee floor), but Sharpe was only 0.53 (CI [-0.093, 1.22]) with 20.9% max drawdown and ~-38% end-unrealized on losers held to the cap. It was profitable in every year including the 2022 crash, with low beta (0.065) and positive alpha (0.063), confirming genuine credited carry rather than disguised long beta.