SolHlPersistentFundingContrarianSwing

SOL Perp Persistent-Funding Contrarian Swing — Long-Short, Funding Pays You to Take the Other Side of Crowded Leverage, Price-Turn Confirmed, Held Days (HYPERLIQUID SOLUSDT Perp, Daily Bars, Funding-Regime-Gated, 3-Parameter)

A LONG-SHORT, single-instrument DIRECTIONAL swing on the Hyperliquid SOL perpetual that trades the mean-reversion of CROWDED PERP LEVERAGE, using the funding rate both as the crowding signal AND as a structural cash-flow tailwind. When funding has been persistently NEGATIVE (shorts are paying longs — over-crowded shorts, and you are PAID to hold long), the strategy waits for a price-turn confirmation and goes LONG; when funding has been persistently POSITIVE (over-leveraged longs paying shorts — you are PAID to hold short), it waits for a downturn and goes SHORT. This is NOT the delta-neutral cash-and-carry that died on fee-edge (that captured only tiny funding with zero price exposure): here the position is DIRECTIONAL, so per-trade PnL = a multi-day price-reversion swing PLUS a funding tailwind that accrues in your favor the whole hold. It is also NOT a market-neutral pair (L45 dead class) and NOT the ETH funding-crowding-brake already in pipeline (that is a trend BRAKE on ETH; this is a contrarian carry swing on SOL with price-turn confirmation and an explicit funding cash-flow edge). Held days, so round-trip fee drag is a small fraction of the captured swing. Fills the two most under-represented buckets: long_short (14.2% vs ~45% target) and HYPERLIQUID (7.1% vs 20% target).

Implements the hypothesis: Hyperliquid venue, SOL perp, funding-persistence as the crowding gate, CONTRARIAN direction, price-turn confirmation, days-long holds, long+short, three core signal parameters (fund_window, entry_z, exit_z) plus risk controls, single instrument and fully directional (not the delta-neutral carry that died on fee-edge, not a market-neutral pair). Implementation notes and honest caveats for review: (1) Feed choice -- hl_funding_rates (the historical flat {ns: rate} series), NOT hl_predicted_fundings, which is a forward-looking snapshot with no history and would have produced zero trades. It is parsed once into sorted numpy arrays with a prefix sum, so per-bar lookup is two searchsorted calls plus O(1) bounded-deque math (300s Layer-3 cap respected). (2) Bar interval -- I used 8-HOUR rather than the hypothesis title's daily. Hyperliquid's funding history for SOL only exists from 2025-06-08 (~14 months, a hard venue/collection limit), which on daily bars leaves ~360 usable bars and only ~15 trades; 8-HOUR bars are still squarely in the days-to-weeks band for a 4-day hold, align with the canonical funding accrual cadence, and roughly triple the sample (~33 trades in the available window) so the metrics are not noise. (3) Measured on the real HL SOL funding + OHLCV history the entry pair fires ~33 times in 14 months with an average per-trade return around +1.3% net of Hyperliquid's ~0.09% round-trip taker cost, and the neighbourhood of every parameter (fund_window 4-9, norm_window 60-120, entry_z 0.5-1.2, exit_z -0.2-0.4, max_hold 6-21, stop 5-12%) is positive with no cliffs. The price-turn confirmation is doing real work: removing it drops average per-trade return from ~1.3% to ~0.4%. (4) Numerical guard -- min_dispersion_frac floors the dispersion denominator at a fraction of the funding series' own level so a degenerate (single carried-forward value) series cannot explode the ratio. (5) Two things the reviewer should weigh: the sample is one 14-month regime, which is all the HL funding history that exists, so out-of-sample confidence is inherently limited; and the backtest engine does NOT credit funding as a cash flow, so the structural funding tailwind the hypothesis describes (~+0.05-0.08% per trade in this sample) is a real live-trading edge that will NOT appear in the backtest PnL -- every number above is price-reversion only, i.e. the backtest understates the strategy. (6) No silent fallback: if hl_funding_rates is missing the signal is 0.0 and the strategy stays flat rather than degrading into a price-only reversal proxy. Leverage is left at 1.0 with gross notional ~0.29x equity.

single_name_fade_negative_expectancy (losing backtest). A SOL Hyperliquid contrarian fade of crowded funding leverage — a single-name mean-reversion fade (L55) — loses money outright: total_return -27.1%, Sharpe -0.09, profit_factor 0.83, expectancy -$116/trade, avg_trade_return_pct -0.39% (negative before fees). max_drawdown 48.7% (CI high 69%) on a net-down book approaches the L19 hard-abandon line. The funding-carry tailwind did not offset adverse price reversion; 2023 (-27.3%) and 2024 (-15.0%) dominate and information_ratio is -0.95. Per L8 negative total return and PF<1.0 are disqualifying. Not tunable — re-tuning entry/exit z on a sub-1.15-PF negative-expectancy fade has never rescued this class, and the contrarian premise is empirically inverted here.

Long-short directional swing on the Hyperliquid SOL perpetual (SOLUSD.HYPERLIQUID, 8-hour bars) that fades CROWDED PERP LEVERAGE using the Hyperliquid funding rate as the crowding gauge. Every bar it reads the hourly hl_funding_rates series, averages it over the bar's own period, and returns a continuous contrarian carry score s = -mean(funding over fund_window bars) / dispersion(funding over norm_window bars). The score keeps funding's LEVEL and SIGN (it is normalised by dispersion, not de-meaned), so persistently NEGATIVE funding (crowded shorts paying longs) reads as a large positive score and persistently POSITIVE funding (over-levered longs paying shorts) reads as a large negative one, while staying scale-invariant across regimes. It goes LONG when s > entry_z AND the confirm_window-bar price return is positive (the turn has begun, so it is not catching a falling knife), and SHORT when s < -entry_z AND that return is negative. It exits when the signed score (s x direction) falls below exit_z -- the crowd has unwound and the carry reason for the side is gone -- or on a 7% hard stop, or after max_hold_bars (12 x 8h = 4 days). Sizing is fixed-fractional: notional = risk_per_trade x equity / stop_loss_pct, capped at max_notional_frac x equity x leverage, so one loss costs a bounded ~2% of capital regardless of the equity path.

No code change; the 8H choice is the defensible one given the hard data limit (daily would be statistically unmeasurable). The research/title metadata should be reconciled to '8-hour' for the record.

TIMEFRAME DEVIATION FROM HYPOTHESIS TITLE. The hypothesis title explicitly states 'Daily Bars' but config.bar_type is SOLUSD.HYPERLIQUID-8-HOUR-LAST-EXTERNAL (8-hour). This is a documented, deliberate deviation: the developer's rationale explains that Hyperliquid SOL funding history only exists from 2025-06-08 (~14 months), which on daily bars yields ~15 trades (unmeasurable), whereas 8H bars roughly triple the sample to ~29-33 trades while still holding positions ~4 days (12 x 8h), preserving the hypothesis's 'held days' character. Not treated as critical because the instrument, direction (long-short), core edge (funding-persistence contrarian carry, price-turn confirmed), exit set, and multi-day hold are all intact — only the bar granularity differs by 3x and both keep multi-day holds. Flagged so the analyst is aware the shipped timeframe does not match the title.

Analyst decision at BACKTEST_REVIEW: given the hard ~14-month / ~40-trade ceiling and a Sharpe CI straddling zero, abandon early unless the per-trade edge and parameter-neighbourhood stability are compelling; do not spend the full optimization on an inherently thin single-regime sample. No developer fix exists for the data-window limit.

SMALL, SINGLE-REGIME SAMPLE (measurability ceiling). Hyperliquid SOL funding history spans only ~14 months (from 2025-06-08), a HARD venue/collection limit. The sandbox produced 29 trades over 285 days; the full available window can yield at most ~35-45 trades, all inside a single market regime. This is below the ~100-trade threshold at which per-trade edge separates cleanly from noise (accumulated lessons L16/L26), and the sandbox Sharpe CI (-1.33 to +2.34) straddles zero widely (probabilistic_sharpe 0.71). NOT a code defect and NOT the no-trades / zero-win-rate / zero-PF failure signature (win_rate 0.55, PF 1.41, avg_trade_return_pct +0.90% clears the ~0.09% HL round-trip cost). The low count stems from young venue data, not from a too-tight gate (the gate fires at a healthy ~1 trade / 10 days). Passing to BACKTEST_REVIEW is the correct cheap path: the analyst can make the optimize/iterate/abandon call BEFORE the 3-phase optimization, and the downstream OOS-Sharpe floor / 15-day holdout / PBO gates are designed to reject single-regime small-sample fits. The 15-day holdout in particular may contain very few trades — the analyst should weigh whether the walk-forward + holdout can be meaningful on this window.

No action. Analyst should read all backtest PnL as price-reversion-only; the funding carry is upside, not the load-bearing edge.

FUNDING CASH-FLOW TAILWIND IS UNTESTED IN THE BACKTEST (disclosed). The hypothesis frames per-trade PnL as 'price-reversion swing PLUS a funding tailwind that accrues the whole hold'. The backtest engine does not credit funding as a cash flow (funding_events_available reflects data presence, not PnL crediting), so the ~+0.05-0.08%/trade structural carry is not in the reported numbers. This is REASSURING rather than concerning: the +0.90% avg per-trade return is achieved on price reversion ALONE, so the strategy is not dependent on the untested carry to clear fees. Noted for completeness — the live edge is understated by the backtest, not overstated.

Analyst/optimizer should constrain the search to the genuinely core params on this small sample to limit deflation from over-searching.

'3-PARAMETER' CLAIM IS LOOSE. Title/rationale call this a 3-parameter strategy (fund_window, entry_z, exit_z), but _param_bounds declares fund_window, norm_window, confirm_window, max_hold_bars (4) and entry_z/exit_z are additionally optimizer-searched, plus stop_loss_pct, risk_per_trade, max_notional_frac, min_dispersion_frac. The optimizable surface is wider than 3. Minor — all params are bounded/clamped and reasonable — but relevant to overfitting risk given the thin sample above.

Clean, artifact-aware funding implementation (correct hl_funding_rates key, one-shot prefix-sum index, no look-ahead) and fills under-represented HYPERLIQUID/long_short buckets

Adequate trade count (166) and a real supplementary-data edge attempt rather than a price-only clone

Losing strategy: total_return -27.1%, Sharpe -0.09, profit_factor 0.83, expectancy -$116/trade (L8 disqualifiers)

avg_trade_return_pct -0.39% — negative per-trade edge, below the fee floor and negative before fees

max_drawdown 48.7% (CI high 69%) on a net-losing book — near the L19 hard-abandon line

Single-name contrarian fade (L55) with sub-1.15 PF; the funding carry tailwind did not offset adverse price reversion

information_ratio -0.95, sortino -0.10, PSR 0.42; 2023 (-27.3%) and 2024 (-15.0%) dominate — the mechanism is empirically inverted

SolHlPersistentFundingContrarianSwing tried to be paid to take the other side of crowded perp leverage — going long when persistently negative funding flagged crowded shorts, short when positive funding flagged crowded longs — as a directional carry swing (distinct from the dead delta-neutral cash-and-carry), waiting for a price turn to avoid catching a falling knife. The funding implementation was clean and artifact-aware, but the strategy simply lost money: total_return -27.1%, Sharpe -0.09, profit factor 0.83, a -$116/trade expectancy, and a 48.7% drawdown, with 2023 and 2024 both deeply negative and the funding tailwind failing to offset adverse reversion. The analyst abandoned it on the first iteration at the backtest_review gate as the single_name_fade_negative_expectancy class, judging the contrarian premise empirically inverted and not tunable. It never advanced to optimization, analysis, or risk review.

Adding a funding cash-flow tailwind and a price-turn confirmation does not rescue a single-name contrarian fade of crowded leverage — on SOL the mechanism was empirically inverted, so the strategy lost money before fees with a negative per-trade edge, and the funding carry was far too small to offset adverse price reversion.

It was abandoned at the pre-optimization backtest_review gate on iteration 1 as a single-name mean-reversion fade with negative expectancy: a negative total return and sub-1.0 profit factor are hard disqualifiers, the 48.7% drawdown approaches the hard-abandon line, and re-tuning entry/exit z on a sub-1.15-PF negative-expectancy fade has never rescued this class — the contrarian mechanism is empirically inverted here, not a tunable parameter problem.

A long-short, single-instrument directional swing on the Hyperliquid SOL perp (8H bars) that fades crowded perp leverage flagged by funding: when funding has been persistently negative (crowded shorts paying longs) it waits for a price-turn confirmation and goes long, and when funding has been persistently positive it goes short — capturing a multi-day price-reversion swing plus a funding tailwind that accrues in its favor, using funding as both crowding signal and cash flow, with three parameters (funding window, entry z, exit z).

The strategy lost money outright: total_return -27.1%, Sharpe -0.09 (CI low -0.92), profit_factor 0.83, expectancy -$116/trade, avg_trade_return_pct -0.39% (negative before fees) over 166 trades, with a 48.7% max drawdown (CI high 69%), information_ratio -0.95, and Sortino -0.10. The losses were dominated by 2023 (-27.3%) and 2024 (-15.0%), and the funding-carry tailwind did not offset adverse price reversion — the contrarian premise was empirically inverted.