XrpFundingExtremeNegativeReversionLongDaily

XRP Funding Rate Extreme Negative Reversion Long on Daily Bars

A long-only single-instrument contrarian strategy on XRPUSDT perpetual futures using daily (1D) bars and the funding rate as the primary signal source. The thesis is asymmetric and microstructure-driven: when 8-hour funding becomes deeply negative (shorts paying longs heavily), it signals overcrowded short positioning that is structurally vulnerable to a short squeeze. XRP is chosen specifically because it is one of the most retail-driven, narrative-sensitive perpetuals on Binance (regulatory news flow, ETF speculation, payments narrative) yet has deep enough liquidity ($500M+ open interest) to trade meaningfully. Critically, this strategy is differentiated from all 10 failed OHLCV-only patterns (none used funding rates), from SolFundingContrarianShort (opposite side: that one shorts on positive funding; this longs on negative funding), and from all other portfolio entries (no pairs, no liquidation cascade, no ratio trade). The strategy is intentionally simple — only 4 parameters — to avoid the overfitting that likely caused verification timeouts in prior multi-filter momentum/mean-reversion attempts. Uses funding_rate_updates supplementary data plus daily OHLCV; no order book, no ticks, no multi-timeframe — keeps verification deterministic and fast.

Iteration 6. I investigated the reported Layer-2 failure rather than assuming it. The `_bar_ts()` TypeError is a real proxy artifact: the Layer-2 proxy rebinds every callable class attribute through types.MethodType, and a base @staticmethod fetched off an instance has no __func__, so it is bound as an instance method and its `bar` slot receives `self`. I reproduced that in isolation. It only fires for code that CALLS `self._bar_ts(bar)` while inheriting the base staticmethod — which is v2/v3 of this strategy, not previous_code. Running the real StrategyVerifier against the actual on-disk v4, v5 and v6 files, all six synthetic scenarios pass on every one, so the iteration_feedback was stale (carried over from the v3 run). previous_code was already immune on two axes (never calls base _bar_ts; overrides it arity-tolerantly), and I preserved both, documenting the verified root cause so it is not 'fixed' again. The change that actually matters is a forward-looking one I found while de-risking Layer 3: backtest_agent.py:45 uses a 365-day window, and at the old entry_funding=0.0003 XRP's trailing-3 funding average crosses that tail exactly ONCE in the last 365 days. Simulating the full entry/exit logic over real XRPUSDT funding history gives 1 distinct trade at 0.0003 versus 12 at 0.0001 (730d: 1 vs 18). A 1-trade backtest would clear the >=1-trade gate and then be abandoned as statistically meaningless. So I loosened the single default to 0.0001, which is ~the 5th percentile of XRP's realized funding distribution (median +0.0001, p5 -0.000114, p1 -0.000296) — still a genuine crowded-short tail, not a regime shift. This stays fee-viable because the edge is the directional squeeze, not funding carry: the 10% take-profit and 6% stop dwarf the ~0.10% round-trip taker cost, and the threshold remains a tunable parameter for the optimizer to explore. Venue is BINANCE USD-M futures because the edge is funding-rate-based and requires holding the perp; leverage stays 1.0 and is deliberately not referenced in sizing, so there is no leverage_set_but_unused violation. I confirmed the exact submitted code passes Layer 1 (static) and all six Layer 2 scenarios with zero exceptions and no frozen-signal warnings — no earlier passing layer was regressed.

failed deflated Sharpe: DSR=0.0488, expected-max=3.756 over 225 trials — the selected optimized Sharpe (1.9527) sits BELOW the best-of-N luck bar and is statistically indistinguishable from noise (is_significant=false, sharpe_ci_low=-0.5161 straddling 0, PBO=0.7381>0.5; probabilistic_sharpe 0.9636 vs DSR 0.0488 is the textbook PSR-vs-DSR selection-inflation trap). The walk-forward's is_overfitted=false is SPURIOUS: out_of_sample_sharpes are [1.848, 5.710, 0.0] — the most recent OOS window produces ZERO trades and the 2.519 average is carried by one implausible 5.71 window; I predicted this empty window at the pre-optimization gate from the 17-month trade gap (2023-09-14 -> 2025-02-28) and it materialized. The untouched holdout also FAILED and is NEGATIVE (holdout_sharpe -0.1202, ratio -0.048), though as a final-15-day slice on a ~10-trades/year strategy it corroborates rather than drives the verdict. Sensitivity FAILED with 3 cliffs (stop_pct, entry_funding, funding_avg_n), and best_params ESCAPED THE SWEPT GRID on both flagged signal cliffs: funding_avg_n=1 (grid tested only {2,3,4}) and entry_funding=0.0001445 (grid topped at 0.00012) — parking on the high side of cliffs it was never measured on. funding_avg_n=1 also destroys the premise, replacing the hypothesis's '~1 day of 8h prints' sustained-positioning average with a single spot print. Fundamentally the strategy adds no risk-adjusted value: NEGATIVE information_ratio in both base (-0.5775) and optimized (-0.5725) forms with benchmark_meaningful=true, so it loses to holding XRP despite near-zero beta (0.0006). The track is outlier-carried — 3 of 62 trades produce 69.8% of total PnL (kurtosis 22.26, skew 2.13) — and the edge is dead in recent regimes: the 20 trades of 2025-2026 contribute just $201 (2.4% of PnL), with optimized 2023 at -0.38%, 2026 at -1.19%, and rolling Sharpe turning negative (-0.229) in March 2026. Costs are NOT the problem (commission 3.41% of gross, impact 1.79%, capacity $312.9M, mean per-trade return 1.49%); significance and forward-generalization are. Not iterate (attempt 1 of 2): the ENTIRE sensitivity surface spans Sharpe 0.03-1.28, uniformly far below the 3.756 luck bar, so there is no robust region above significance to tune toward and a second best-of-225 sweep re-selects noise; the optimizer's flattering refit (Sharpe 0.52 -> 1.95, max_dd 7.36% -> 2.20%) was bought purely by halving exposure (avg_position_pct 24.84 -> 8.97), which shrinks drawdown without creating edge. Not revise_hypothesis: XRPUSDT.BINANCE Daily is a valid, data-rich instrument with verified full funding coverage (7,013 rows, 2020-01-06 -> 2026-05-31) and an adequate 64-trade base sample, so per the overfit->abandon rule this is overfitting-to-search on a fine target, not a proven mechanism stranded on a dead one. FAILURE PATTERN: a deeply-negative-funding crowded-short reversion long on single-asset XRP daily has a REAL but regime-bound carry+squeeze edge that concentrates almost entirely in the 2022 persistent-negative-funding bear market (23 of 62 trades; 3 trades = 69.8% of PnL) and decays to ~zero by 2025-2026 ($201 across 20 trades). Its ~10-trades/year cadence is too sparse to survive per-window walk-forward (one OOS window empty) or best-of-225 selection: DSR 0.0488 with the optimized Sharpe 1.95 below a 3.76 luck bar, PBO 0.738, a negative holdout, 3 cliffs whose optimum lies OUTSIDE the swept grid, and a negative information ratio vs its own XRP buy-hold. A non-overfit-flagged walk-forward (IS/OOS ratio 1.21), a high probabilistic_sharpe (0.9636), a doubled optimized Sharpe, and a fat per-trade margin (1.49% vs 0.10% costs) measure consistency and cost-survivability — NOT statistical significance or forward-generalization. Two structural tells to reuse: (1) an avg-OOS carried by one window while another is empty must never be credited as passing, and (2) when best_params lands outside the sensitivity grid on a parameter already flagged is_cliff, the selected config is unmeasured by definition.

Long-only contrarian strategy on XRPUSDT.BINANCE USD-M perpetual, daily bars. Uses the 8-hour funding rate as the positioning signal: when the trailing average of the last 3 funding prints (~1 day) turns deeply negative (shorts paying longs heavily), it reads as overcrowded short positioning that is structurally vulnerable to a squeeze, so the strategy goes long. calculate_signal returns -avg_funding in natural funding units every bar (continuous, not gated); should_enter thresholds it against entry_funding in matching units. Exits on any of: funding normalizing back above exit_funding (squeeze resolved), a 10% take-profit, a 6% stop-loss, or a 7-day max-hold time stop. Position sizing is capital-relative via get_account_equity(), capped by both a notional fraction (30%) and a risk-per-trade budget (1.5% equity risked against the 6% stop). If funding data is absent the strategy stays flat and returns a tiny clipped price-ROC that can never cross the entry threshold, so there is no silent price-only fallback edge.

Verification failed (Layer 2 — synthetic scenarios): Parameters used: ['tp_pct', 'risk_pct', 'stop_pct', 'exit_funding', 'min_notional', 'notional_pct', 'entry_funding', 'funding_avg_n', 'max_hold_bars', 'lag_tolerance_s'] Check that __init__ sets all attributes from self.parameters.get(). - steady_uptrend: TypeError: FactoryStrategy._bar_ts() takes 1 positional argument but 2 were given (bar timestamp: 1735689660000) - steady_downtrend: TypeError: FactoryStrategy._bar_ts() takes 1 positional argument but 2 were given (bar timestamp: 1735689660000) - flat_ranging: TypeError: FactoryStrategy._bar_ts() takes 1 positional argument but 2 were given (bar timestamp: 1735689660000) - volatility_spike: TypeError: FactoryStrategy._bar_ts() takes 1 positional argument but 2 were given (bar timestamp: 1735689660000) - zero_volume: TypeError: FactoryStrategy._bar_ts() takes 1 positional argument but 2 were given (bar timestamp: 1735689660000) - price_gap: TypeError: FactoryStrategy._bar_ts() takes 1 positional argument but 2 were given (bar timestamp: 1735689660000)

Trades match the hypothesis exactly: 64/64 LONG, zero shorts, and funding income of +$1,299.88 (positive on 50 of 64 trades) confirms entries genuinely land on negative-funding/crowded-short bars. This is a verified mechanism match, not a coincidental edge.

Funding data has FULL coverage: data/supplementary/fundingRate/XRPUSDT/all.parquet holds 7,013 rows spanning 2020-01-06 -> 2026-05-31 across the entire 2,374-bar backtest. No silent truncation.

Per-trade edge clears trading costs with margin: mean pnl_pct 0.395%, median 0.563%, against ~0.10% round-trip taker cost on BINANCE USD-M. commission_pct_of_gross is only 4.19% and impact_cost_pct 7.86% — fees are NOT what is limiting this strategy.

Sample size is adequate for a 4-parameter sweep: 64 trades over 6.5 years, entry_diag shows entry_signaled 64 == entries_submitted 64 with zero drops for sizing, notional, or cash.

Zero trades in 2024 is correct behavior, not a defect — funding was persistently positive through that bull run, so a negative-funding trigger should stay flat. The signal is doing what it claims.

Funding is credited by the engine and contributes ~23% of total profit, so the carry component is real and measurable rather than invisible.

Severe single-regime concentration: 2022 returns +13.27% while the FULL-RUN total_return is only +4.41%. Annual returns are 2020 -5.25%, 2021 -1.54%, 2022 +13.27%, 2023 +0.45%, 2024 absent, 2025 -3.05%, 2026 +2.71% — five of six years are flat-to-negative. 26 of 64 trades sit in 2022, the persistent-negative-funding bear regime.

A 17-month trade gap (2023-09-14 -> 2025-02-28) means at least one walk-forward OOS window will very likely contain ZERO or near-zero trades, which historically produces OOS Sharpe 0.0 and an is_overfitted=TRUE flag.

Risk-adjusted metrics are marginal at base: sharpe_ratio 0.518 with sharpe_ci_low -1.7199 straddling zero, sortino 0.495 BELOW Sharpe, profit_factor 1.175 (under the 1.2 guideline), and probabilistic_sharpe 0.686.

information_ratio is NEGATIVE (-0.5775) with benchmark_meaningful=true — it delivers less risk-adjusted value than simply holding XRP, despite near-zero beta (0.0061) and negligible alpha (0.002).

Fat left tail: tail_ratio 0.7565 (<1), return_kurtosis 12.65, cvar_95 2.16% vs var_95 1.41%, avg_loss ($1,161) exceeds avg_win ($1,062). max_drawdown_duration_days is 905 and the most recent rolling_sharpe reading is -0.0082.

The daily-bar stop is porous: trade 1 shows mae_pct -35.98% against a 6% stop_pct because stops only evaluate at bar close (realized -7.39%, largest_loss -$3,632 on ~$25k notional). Not a bug, but a real tail risk the backtest understates.

Mild pre-fitting already occurred: the developer's docstring records lowering entry_funding from 0.0003 to 0.0001 because the original default fired only once. The optimizer will sweep this parameter, so it is disclosed rather than hidden, but the default is not naive.

64 LONG / 0 SHORT, +$1300 funding received

long-only on negative funding

The implementation is correct and the mechanism is faithfully expressed: 62/62 LONG, entries on negative funding, funding credited as income. Data coverage is complete (7,013 funding rows, 2020-01-06 -> 2026-05-31), verified pre-optimization.

Fees are not the binding constraint: commission_pct_of_gross 3.41% optimized (4.19% base), impact_cost_pct 1.79%, capacity_usd $312.9M.

The account never liquidated, and mean per-trade return (1.49%) comfortably clears the ~0.10% round-trip futures cost. The failure is statistical, not economic.

walk_forward exit-path and trade-count behavior are internally consistent; no engine or sizing anomaly contaminates this result.

DECISIVE — failed the multiple-testing gate: deflated_sharpe 0.0488 against a 0.95 bar, with the selected optimized Sharpe (1.9527) BELOW its own 225-trial expected-max luck bar of 3.756. is_significant=false, sharpe_ci_low=-0.5161 straddling zero, pbo=0.7381 (a 74% probability of backtest overfitting). probabilistic_sharpe 0.9636 vs deflated_sharpe 0.0488 is the textbook PSR-vs-DSR selection-inflation trap.

The walk-forward's is_overfitted=false is SPURIOUS. out_of_sample_sharpes are [1.8479, 5.7102, 0.0]: the most recent window produces ZERO trades, and the 2.519 average is carried entirely by a single implausible 5.71 window. I predicted this empty window pre-optimization from the 17-month trade gap (2023-09-14 -> 2025-02-28); it materialized exactly.

Sensitivity FAILED with 3 cliff parameters (stop_pct, entry_funding, funding_avg_n), and best_params escaped the swept grid on BOTH flagged signal cliffs: funding_avg_n=1 when the grid tested only {2,3,4}, and entry_funding=0.0001445 when the grid topped out at 0.00012. The optimizer parked on the high side of two cliffs it was never measured on.

funding_avg_n=1 abandons the stated premise. The hypothesis specifies averaging '~1 day of 8h prints' to detect SUSTAINED crowded-short positioning; a single 8h print is a spot reading. The optimizer traded the mechanism away for Sharpe.

No robust region exists to tune toward: the ENTIRE sensitivity surface spans Sharpe 0.03-1.28, uniformly far below the 3.756 luck bar.

NEGATIVE information_ratio in BOTH forms — base -0.5775, optimized -0.5725 — with benchmark_meaningful=true. Despite near-zero beta (0.0006) it delivers less risk-adjusted value than simply holding XRP.

Severe outlier concentration confirmed from the detail file: 3 of 62 trades produce 69.8% of total PnL ($2,798 + $1,780 + $1,379 of $8,537). return_kurtosis 22.26, return_skew 2.13.

The edge is dead in recent regimes: the 20 trades in 2025-2026 contribute $201, just 2.4% of total PnL. Optimized annual returns are 2023 -0.38% and 2026 -1.19%, and optimized rolling_sharpe goes negative (-0.2294 on 2026-03-05), ending at 0.848.

The optimizer's flattering refit (Sharpe 0.518 -> 1.953, max_drawdown 7.36% -> 2.20%) was bought by HALVING exposure (avg_position_pct 24.84 -> 8.97, exposure_pct 6.69 -> 4.92), which shrinks drawdown without creating an edge. max_drawdown_duration_days remains 805.

The holdout also failed and is negative (holdout_sharpe -0.1202, ratio -0.048), though I weight this lightly: a final-15-day holdout on a ~10-trades/year strategy is thin evidence on its own. It corroborates rather than drives the verdict.

Do NOT send this to optimization — 12 trades and 1.11% exposure will produce empty walk-forward OOS windows and a guaranteed deflated-Sharpe failure. Fix the sample first, in this order: 1. DIAGNOSE THE FUNDING FEED BEFORE CHANGING ANY THRESHOLD. In `_ensure_funding()`, after building `self._fund_ts`, log `len(self._fund_ts)`, `min(self._fund_ts)`, `max(self._fund_ts)`, and the median gap between consecutive prints. The backtest reports `funding_events_available = 111` for XRPUSDT over 2020-2026 while sibling BTC/BNB strategies report 1,421-3,268 — if your parsed index really is ~111 prints spanning ~37 days, the signal is undefined on ~98% of bars and NO parameter change will help. Report the actual number back. Do not assume the cause; measure it. 2. IF THE PARSED INDEX IS MUCH SMALLER THAN THE RAW DICT, it is a parsing bug, not a data gap. Note that `_norm_ns()` is applied to dict KEYS — for the documented dual-key funding dict `{ns_int: rate, "ns_str": rate}`, `_norm_ns(k)` calls `int(k)` on both forms, which is correct, but confirm no keys are silently dropped by the bare `except Exception: continue`. Add a counter for skipped keys and log it. 3. IF THE FEED IS GENUINELY SPARSE, the XRPUSDT-daily target cannot test this hypothesis. Say so explicitly in your response rather than loosening thresholds to manufacture trades on 37 days of data — that would be fitting to whichever handful of days happen to have coverage. 4. ONLY IF THE FEED IS DENSE (thousands of prints) is the threshold the binding constraint. In that case replace the absolute `entry_funding = 0.0003` with a rolling-percentile trigger (e.g. current 3-print average in the bottom 10-15% of its trailing 90-day funding distribution, with a `min_funding_samples` guard), which self-calibrates as XRP's funding regime shifts. Target >= 60 trades. Be aware a sibling strategy (BnbDailyFundingPercentileLong200Sma) showed the percentile version still decays post-2021 — so also report per-year trade counts so we can see whether the signal survives 2023-2026 or is a 2020-2022 artifact. 5. Also raise `funding_avg_n` exposure to scrutiny: at 3 prints it is a 1-day average, so `lag_tolerance_s = 86400` (24h) can admit a window whose oldest print is ~2 days stale. Tighten `lag_tolerance_s` to ~32400 (9h, one cadence + slack) so the 'recent funding' the signal reads is actually recent. Re-run the backtest and report: parsed funding print count + span, trades per calendar year, and exposure_pct.

XrpFundingExtremeNegativeReversionLongDaily bet that deeply negative XRP funding marks overcrowded shorts and traded long into the squeeze, using only four parameters and full funding coverage (7,013 rows, 2020-01-06 to 2026-05-31). The mechanism was real and cheap to run — 64/64 trades were long, funding was credited as income, commissions were just 4.19% of gross, and per-trade returns cleared costs with margin — so the backtest-review gate cleared it to optimize despite marginal Sharpe (0.518) and a negative information ratio. But the edge was regime-bound: it concentrated in the 2022 persistent-negative-funding bear market, 3 of 62 trades produced 69.8% of PnL, and the 20 trades of 2025-2026 contributed just $201. After 6 iterations the analyst abandoned it on statistical significance rather than economics — deflated_sharpe 0.0488, an optimized Sharpe below its own best-of-225 luck bar, PBO 0.738, a negative holdout, and no robust region anywhere on the sensitivity surface to tune toward.

An average out-of-sample Sharpe carried by one strong window while another window is empty must never be credited as passing, and when best_params lands outside the sensitivity grid on a parameter already flagged is_cliff, the selection is untested at its own optimum.

The analyst returned an `abandon` verdict after optimization: deflated_sharpe was 0.0488 against a 0.95 bar, the selected optimized Sharpe (1.9527) sat below its own 225-trial expected-max luck bar of 3.756, PBO was 0.7381, the holdout failed and was negative (holdout_sharpe -0.1202, ratio -0.048), and sensitivity failed with 3 cliff parameters (stop_pct, entry_funding, funding_avg_n) whose best_params escaped the swept grid. The pre-optimization backtest_review had passed it with `optimize` while explicitly flagging the 17-month trade gap that later produced a zero-trade OOS window.

Go long XRPUSDT perpetual futures on daily bars when 8-hour funding turned deeply negative, on the thesis that shorts paying longs heavily signals overcrowded short positioning vulnerable to a squeeze.

The base backtest produced 64 long trades (zero shorts) over 218 data days with total_return 4.41%, Sharpe 0.5177, win rate 56.3%, profit factor 1.175, max drawdown 7.36%, and a negative information ratio of -0.5775; optimization refit it to Sharpe 1.9527, total_return 8.16%, and max drawdown 2.20% across 62 trades, but only by halving exposure (avg_position_pct 24.84 → 8.97).