AvaxHlFundingWashoutReversalLS

Funding-Positioning Washout Reversal, Long-Short (AVAXUSD.HYPERLIQUID, 4H bars) — take a DIRECTIONAL contrarian position after funding z-score reaches a crowding extreme, ride the multi-day positioning unwind, 2-parameter

A LONG-SHORT, single-instrument, multi-day-hold DIRECTIONAL reversal strategy on the AVAX Hyperliquid perpetual (AVAXUSD.HYPERLIQUID), 4H bars, whose signal is the perpetual FUNDING RATE relative to its own trailing distribution. This is deliberately NOT any of the dead funding classes: it does NOT hold a delta-neutral spot/perp leg to harvest the funding cashflow (that is L87 cash-and-carry, which dies fee_edge because collected funding < fees + basis drift), and it does NOT fade the mark-vs-index premium (L85). The edge here is the PRICE move, not the cashflow: when funding pushes to a positive extreme, perpetual longs are over-leveraged and crowded, and the subsequent forced/voluntary de-risking produces a multi-day price PULLBACK that is large relative to fees; symmetrically, a negative funding extreme marks crowded shorts that squeeze upward. We take a directional position AGAINST the crowd and hold for the positioning unwind (1-3 days), not a scalp. Chosen to fill three quota gaps at once: HYPERLIQUID venue (7.2%, under the 20% target), long_short direction (14.4%, under the 55% cap and heavily under-represented), and a funding-driven mechanism that is directional rather than the exhausted hedged-carry framing. Single-instrument by design to avoid the HL multi-instrument basket backtest_failed trap (L88). 4H bars over Hyperliquid's ~2.3y of usable 4H history (~5000 bars) is enough to validate without hitting the HL sub-hourly data wall.

Iteration 2 addresses the backtest gate's finding that the realised equity curve (-10.12%) disagrees in sign with the mtm-derived Sharpe (0.72): judged on the equity curve the strategy loses, because entries fired on the FIRST bar funding reached an extreme, while the crowd was still building — the position then absorbed the last leg of the trend, producing large adverse excursions that hit the stop before the unwind paid (mark-to-market averaged out positive across bars, realised round-trips did not). The smallest change that fixes this is an entry-TIMING one, not a new gate: entries now additionally require the funding z to have TURNED back from a peak/trough reached within the last confirm_bars (24h) — signal computation, exits, sizing, imports and structure are untouched, so Layers 1-3 stay green. Trade count is preserved because every crowding extreme eventually rolls over; only the entry bar moves. The bracket is also made positively skewed (3% stop vs 7.5% take-profit) so the realised curve, not just the mtm series, carries the edge. Signal remains continuous (-z every bar) and is thresholded in should_enter against the matching-unit entry_z.

analyzing_stuck_3_timeouts

Directional contrarian reversal on the AVAX Hyperliquid perpetual (4H bars). The signal is the z-score of 8h-smoothed HL funding versus its own trailing 7-day distribution (point-in-time, staleness-guarded). A positive funding extreme marks crowded, over-leveraged longs; a negative extreme marks crowded shorts. The strategy takes a DIRECTIONAL position against the crowd once the extreme starts to roll over and holds it through the multi-day positioning unwind (min 24h, max 72h), exiting on funding normalisation, a funding flip against the position, the time stop, or the venue-managed bracket (3% stop / 7.5% take-profit, both intra-bar). No hedge leg, no funding cashflow in PnL — the captured return is the price reversion. If the HL funding series is missing or stale the signal is 0 and the strategy stays flat; it never substitutes a price-only proxy.

Analyst to confirm the reversion edge is not concentrated in the calm/stressed tails and that position_frac (0.15) x leverage (2.0) does not overrun the thin HL AVAX book at scale.

Strategy loses in the normal-volatility regime (-20% in sandbox) and impact_cost_pct is 58% of gross with capacity ~$293k. Advisory only — performance is judged at full backtest/optimization, not here.

Sandbox regime split shows the book LOSES in the normal-vol tercile (-20.0% return) while winning in calm (+4.3%) and stressed (+6.6%) — the edge is not stable across regimes and the mtm Sharpe (0.41) has a CI [-0.82, 1.70] straddling zero. Impact cost is 58% of gross with capacity only ~$293k, so the apparent per-trade edge shrinks fast at size. Worth watching on full history whether the funding-washout reversion survives the normal regime and realistic sizing; not a correctness issue.

Numerically viable on a reliable 137-trade / 2.6-year sample (metrics_reliable=true): avg_trade_return_pct 0.646% is comfortably above the Hyperliquid ~0.15% fee floor AND is already net of folded market impact (impact_folded=true), so the per-trade edge survives costs at the tested size.

Trades faithfully implement the stated mechanism: 137 signaled = 137 submitted, balanced long(90)/short(47) directional contrarian entries, venue-managed bracket, single-instrument by design (avoids the HL multi-instrument basket trap). No mechanism mismatch, no artifact signature.

Low headline risk: max_drawdown 9.8%, positive Sharpe 0.414, PF 1.138 > 1.0, recovery_factor 1.45, positive alpha (0.054) with near-zero beta (0.014) — this is genuinely market-neutral directional timing, not closet beta. end_unrealized_pct -1.8% means the return is realized, not an MTM mirage.

Sharpe CI [-0.808, 1.558] straddles zero and PF 1.138 is only marginally above the fee-drag band — the edge is statistically fragile and this is the exact single-perp funding-gated directional class (L73) that repeatedly dies overfit in the 3-phase deflation when PF < ~1.7.

Capacity_usd only ~$280k with impact_cost_pct 59.7% of gross — the edge is real at toy scale but shrinks fast at size; a real promotion blocker later even if optimization succeeds.

Regime instability (QA concern borne out): loses -20.0% in the normal-vol tercile while winning in calm (+3.3%) and stressed (+6.6%); annual returns decaying 8.4% (2024) -> 3.3% (2025) -> 1.9% (2026).

Backtest gate: - metrics unreliable: Incoherent: Sharpe 0.72 disagrees in sign with total_return -10.12% — the Sharpe is computed from the mtm_equity_curve series while the return comes from the equity curve. Judge on the equity curve.

Benjamini-Hochberg at q=0.10 over 261 programme candidates keeps 4. A candidate that does not survive here is not distinguishable from the programme's own noise, however good its individual statistics look.

AvaxHlFundingWashoutReversalLS-85da70ac61

AvaxHlFundingWashoutReversalLS bet that funding-rate crowding extremes on the AVAX Hyperliquid perp would precede multi-day directional price unwinds, taking contrarian long-short positions against the crowd. Its second-iteration backtest looked reasonable on the surface — 137 trades, Sharpe 0.414, profit factor 1.138, and a per-trade return of 0.646% comfortably above trading costs — so the backtest reviewer sent it to optimize. Optimization then dismantled the case: the walk-forward went from strong in-sample Sharpes to a deeply negative -1.59 average out-of-sample, the holdout failed with a negative Sharpe on just 8 trades, and the candidate did not survive deflation or the programme FDR. The strategy was ultimately abandoned when the analyzing stage timed out three times and never produced a verdict, ending a run that the robustness statistics had already marked as overfit.

A per-trade edge that clears the fee floor in-sample is not enough for this slow single-perp funding-gated class: with a fragile Sharpe CI straddling zero and PF near the fee-drag band, it collapsed out-of-sample — future iterations need a robustly positive OOS/holdout signal before optimization, not just a viable raw backtest.

The run was abandoned at the ANALYZING stage after the analyst stalled and timed out three times (abandon_reason "analyzing_stuck_3_timeouts"), so no final analyst verdict was ever recorded — though the optimization evidence (negative OOS Sharpe, failed holdout, overfit=true) already pointed away from promotion.

A long-short, single-instrument directional contrarian reversal on the AVAXUSD Hyperliquid perpetual (4H bars) that used funding-rate z-score extremes as a crowding signal, fading over-leveraged positioning to capture the multi-day price washout rather than any funding cashflow.

The initial 137-trade backtest over ~2.6 years was numerically viable — Sharpe 0.414, total return 12.2%, profit factor 1.138, max drawdown 9.8%, and avg_trade_return_pct 0.646% (above the ~0.15% Hyperliquid fee floor, net of folded impact). But optimization exposed overfitting: walk-forward flipped from an avg in-sample Sharpe of 1.82 to an avg OOS Sharpe of -1.59, the holdout failed (Sharpe -0.208 on 8 trades), PBO was 0.54, deflated Sharpe ~0.002, and it did not survive the programme-level FDR.