ETH Cross-Exchange Funding Differential Arbitrage (BINANCE Perp vs HYPERLIQUID Perp, Market-Neutral, Cross-Venue)

Market-neutral cross-exchange funding-rate arbitrage on ETH: when the funding rate differential between BINANCE USD-M perp and HYPERLIQUID perp exceeds a threshold, simultaneously long the venue with LOWER funding (paying less or receiving more) and short the venue with HIGHER funding (collecting more), capturing the net funding differential as carry while delta-neutral across exchanges. This hypothesis SIMULTANEOUSLY FILLS FOUR severely under-represented portfolio buckets in ONE strategy: (a) HYPERLIQUID venue (4.1% → ≥20% target), (b) cross-venue scope (3.1% → ≥15% target), (c) market-neutral direction (8.8% → reducing 91.2% long-only), (d) pairs scope (12.9% → reducing 80.5% single). The mechanism is COMPLETELY DISTINCT from every refuted mechanism in batch memory (Bull Flag, Pullback-in-Trend, Donchian breakouts, structure patterns, candle patterns) because the edge is a structural funding differential between two independent perp venues, NOT directional price prediction. The analyst has NO refutation for cross-exchange funding arb because it has not been tested in factory record. Even my own parallel BtcSpotPerpFundingCarryHedged hypothesis is mechanically different — that captures absolute funding level via spot+perp; this captures the SPREAD between two perps via long_perp + short_perp on different venues.

Addresses each iteration-2 issue with the smallest change from previous_code, preserving the (passing) signal/data/neutrality scaffolding. (1) STOP THE CHURN: the raw HL leg (rate*8760) flips sign almost every bar, so the annualised differential is now EMA-smoothed (diff_ema_bars=12h) and the trade direction reflects a persistent spread; additionally the `carry < 0` sign-flip exit is GATED behind min_hold_bars (no exit at all during the min-hold window), so positions are held ~7 days and the carry accrues instead of churning out in 1-3h. (2) RAISE THE ENTRY BAR + HYSTERESIS: entry threshold raised to 12% annualised (smoothed) with a wide entry/exit band (exit 4%) and a 48-bar re-entry cooldown so we don't immediately re-open after closing. (3) UNIT ECONOMICS: a new _carry_clears_costs gate only opens a pair when expected carry over the intended min-hold (|D|/100 * min_hold/8760) >= carry_safety * (fee_rt_frac + impact_frac) on BOTH legs (~0.19% RT), so trades are only taken when the carry can beat the fees; at 12% over 168h expected carry (~0.23%) exceeds the 0.19% round-trip fee. (4) NEUTRALITY: both legs sized to the SAME USD notional N (both venues linear), so the ETH exposures cancel and residual directional PnL stays near zero, leaving the funding spread as the sole PnL source; leg_gross_frac reduced to 0.5 to cut market impact. If, after these fixes, held trades still show realized funding < fees, the Binance-vs-HL ETH differential is genuinely too thin and the strategy should be abandoned.

Do NOT abandon — funding is credited correctly (HL legs show non-zero funding in the trades), so the carry mechanism works; the trading logic is destroying it via over-trading. Fixes required before re-backtesting: 1. STOP THE CHURN (root cause). In should_exit, the `carry < 0 -> return True` branch fires before the min_hold_bars check, so positions exit after 1-3 hours instead of the intended ~1 day. The 'carry' here is the annualized differential (HL rate x 8760) which is dominated by hourly HL noise and flips sign almost every bar. Either (a) gate the sign-flip exit behind min_hold_bars (don't exit on a transient flip during the min-hold window), and/or (b) smooth the differential signal (e.g. EMA over several hours) so direction is based on a PERSISTENT differential, not one bar of HL noise. Goal: hold for days, not hours, so the carry actually accrues. 2. RAISE THE ENTRY BAR AND REDUCE FREQUENCY. With the x8760 HL annualization, entry_annual_pct=3.0 is trivially exceeded by noise (359 entries). Require a much larger, persistent differential to enter, and add a wider entry/exit hysteresis band so you don't re-enter immediately after exiting. 3. PROVE THE UNIT ECONOMICS. Each pair costs ~$142+ in commission plus impact (~$67-75 per leg one-way), while funding captured is ~$1-5 per trade. Before re-submitting, instrument a check that the EXPECTED carry over the intended hold exceeds the ~0.19% round-trip fee + impact on BOTH legs. If a sample of held trades does not show realized funding > total fees, the ETH Binance-vs-HL differential is too thin to trade and this should then be abandoned. 4. TIGHTEN NEUTRALITY. Pair-1 shows ~$143 of net directional price PnL (HL long -670 vs Binance short +527) — the legs are not cleanly dollar-neutral. Verify equal USD notional fills on both legs and that residual directional PnL is near zero, so the only PnL source is the funding spread. After these changes, re-backtest; if it still bleeds with funding < fees on held trades, abandon (thin structural edge), but the current result is a fixable over-trading defect, not a dead premise.