BtcHyperliquidBinanceInterExchangeBasisDiscountLong1H

BTC Hyperliquid-Binance Inter-Exchange Basis Discount Long

A long-only, cross-venue strategy that trades BTCUSD.HYPERLIQUID perpetual futures (MARGIN, ~0.09% RT fees) using the BTCUSDT.BINANCE perpetual futures price as a leading microstructure signal. The mechanism: when Hyperliquid trades at a meaningful DISCOUNT to Binance (inter-exchange basis_bps_binance_over_hl >= threshold), this indicates that price discovery has accelerated on Binance (institutional liquidity, larger order book) while HL is lagging. The cheaper venue typically catches up to the richer one within 1-3 days because: (a) cross-exchange arbitrageurs deploy capital to buy HL/sell Binance until convergence; (b) HL's lower fee structure attracts retail flow that follows Binance's directional signal; (c) absent a structural funding-rate divergence (which we filter out), basis mean-reverts naturally. Long-only on HL captures the catch-up move without shorting. Directly builds on the successful BTCSpotPerpPremiumCatchUpLong template (which validated cross-venue basis trading through verification) but ADDRESSES THE MOST UNDER-QUOTA VENUE: HYPERLIQUID is at 6.3% vs 20% target — the largest single quota gap. Also addresses cross-venue (3.0% vs 15%). Distinct from BTCSpotPerpPremiumCatchUpLong because: that strategy uses BTC spot vs BTC Binance perp (intra-Binance venue cross), this uses BTC HL perp vs BTC Binance perp (inter-exchange cross). Different exchange microstructure → different signal source → genuinely additive to the portfolio.

Fixes the named fee-viability defect at its root instead of re-tuning around it. The entry gate is now FEE-ANCHORED in code: entry_threshold_bps is hard-clamped to at least _fee_edge_mult x _fee_round_trip_bps (1.5 x 9 bps = 13.5 bps floor), so no parameter search can ever reproduce iteration 2's sub-fee 2 bps gate; the default is 15 bps and _param_bounds mirrors the clamp. The unreachable 4% take-profit QA flagged is replaced by a 1.0% target scaled to the mechanism's own measured move, exit_threshold_bps is widened to 2.0 and the time stop shortened to 12 bars, which together cut turnover from 173 trades to at most the ~49 qualifying bars in the sample. I checked QA's claim that no fee-clearing threshold can ever trigger against the real catalog: over 7,844 timestamp-aligned HL/Binance 1H bars the basis is centred near zero (median -0.6 bps, p95 +2.1, p99 +7.3), so QA is right that the typical dislocation is sub-fee, but 49 bars (0.62%) do print >= 15 bps. Conditional on that gate the forward HL move averages +36 bps (1h), +53 (4h), +85 (12h) and +114 (24h) versus an unconditional -0.4 / -1.4 / -4.4 / -8.6 bps over the same horizons, and the basis itself decays from +33 bps at entry to ~-1.7 bps within an hour. So the gate clears the venue fee by 1.7x and the expected move by 4-12x. Honest caveat for the analyst: this is a tail trade with roughly two dozen episodes per year, so the walk-forward windows will be thin, and the conditional forward return is measured on 49 non-independent bars; if the backtest shows the per-trade edge is really clustered beta rather than convergence, the family should be abandoned rather than re-tuned again. Everything else -- the SyncedBuffer alignment, the continuous log-return proxy signal, the entry gate plumbing and the sizing -- is unchanged from the code that already passed Layers 1-3.

Thin, tail-gated, recent-only cross-venue basis trade — not worth 2 hours of optimization. The >=15bps HL-vs-Binance discount gate fires only on rare episodes, producing just 28 trades whose PnL all falls in a ~2-month window (2026-07/08) even though the HL leg spans ~11 months — the earlier ~9 months generated almost no qualifying signals, so the 3-phase walk-forward and holdout cannot populate. It is not significant: Sharpe 1.02 with bootstrap CI [-1.28, 2.55] straddling zero, PSR 0.606, effective_daily_obs 25, total_return only +0.78% with exposure_pct 2.24% (capital almost entirely idle). It is also cost-dominated and capacity-dead: commission 24.3% of gross plus impact 15.7% consume ~40% of gross PnL, and capacity_usd is only ~$4.05M. Although avg_trade_return_pct 0.416% clears the HL floor and PF is 3.78, a same-asset BTC perp-perp basis is normally 1-2 bps, so a persistent >15bps 'discount' is as likely a stale-print/bar-misalignment artifact as a real dislocation, and 28 tail trades over one recent regime cannot survive the deflated-Sharpe/holdout gates or scale. This is the L57/L44 cross-venue basis family. Failure pattern: no_edge/capacity cross-venue same-asset basis, thin recent-only tail sample, Sharpe indistinguishable from zero (L57/L44).

Long-only cross-venue catch-up on BTCUSD.HYPERLIQUID 1H perps, using BTCUSDT.BINANCE 1H as a timestamp-aligned signal leg. When Hyperliquid trades at a discount of at least entry_threshold_bps to Binance (default 15 bps, hard-clamped in code to >= 1.5x the ~9 bps HL round trip) and HL is above its own 50-bar SMA, the strategy buys HL and rides the catch-up. It exits when the discount has closed (basis <= exit_threshold_bps), on a 1.0% take-profit sized to the measured catch-up move, a 2.5% stop, or a 12-bar time stop. Sizing is 20% of equity, no leverage.

Verification failed (Layer 4 — QA review) [class=code_defect]: - [edge_concern] Beyond the fee-viability defect: even the directional exits cannot rescue this. A 2-bps entry dislocation implies a ~2-bps catch-up, so the 4% take-profit essentially never fires from the stated mechanism — it would only trigger on unrelated directional BTC moves (beta/luck), not the inter-exchange basis convergence the hypothesis names. The mechanism's gross edge is structurally an order of magnitude below the venue fee, so the research lead should reconsider whether an inter-exchange BTC-perp basis (~1-2 bps) is a viable edge on a ~9-bps round-trip venue at all, rather than re-tuning it. - FEE-VIABILITY FAILURE (checklist item 8). entry_threshold_bps = 2.0 (0.02%) and exit_threshold_bps = 0.5, so the basis-convergence edge harvested is ~1.5 bps = 0.015% gross, against a ~9-bps (0.09%) Hyperliquid round-trip fee — ~6x BELOW the fee floor. Item 8 flags entry_threshold/expected_move < 0.15% as critical; 0.02% is 7.5x below that line. Arithmetic, not a smoke-test artifact: sandbox shows avg_trade_return_pct 0.074% (below the ~0.09% floor), PF 0.76, commission 44.35% of gross over 173 trades, negative total return. The developer moved the threshold the WRONG way for trade count (iter-1 5 bps -> 4 trades; iter-2 2 bps -> churn), and since the basis 'normally sits at ~1-2 bps' per the docstring, no fee-clearing threshold (>~15 bps) can ever trigger. - High turnover with a sub-fee edge: 173 trades over ~328 data_days at avg_holding 2h11m, commission 44.35% of gross — the L6 fee-graveyard pattern, a direct consequence of the sub-fee entry gate.

None required for correctness; the code is internally consistent as a fast-convergence trade.

Horizon compression vs the hypothesis: the hypothesis frames catch-up as '1-3 days', but the convergence exit (basis <= 2 bps) fires in ~1h (sandbox avg holding 1h 0m) with a 12-bar/12h time stop, so it never holds 1-3 days. This is data-calibrated timing of the SAME convergence mechanism (basis decays +33 bps -> ~-1.7 bps within an hour), not a different strategy, so not critical — but the realized trade is a 1-hour basis scalp, which the analyst should read as the actual mechanism.

None — trimming already present.

Static analyzer flagged unbounded growth on self._hl_closes.append(). False positive: the list is trimmed to sma_trend_period + 100 right after the append.

Clean cross-venue implementation with timestamp-aligned legs, fee-anchored entry gate (fixing iteration-2's fee-graveyard churn), and a genuine measured catch-up move; avg_trade_return_pct 0.416% clears the HL floor and PF is 3.78 with a tiny 0.15% drawdown

Thin, tail-gated, recent-only sample: 28 trades and the >=15bps discount only fires on rare episodes — all trade PnL falls in 2026-07/08, so despite ~11 months of HL data the earlier ~9 months produced almost no qualifying signals; walk-forward/holdout cannot populate

Not statistically significant: Sharpe 1.02 with bootstrap CI [-1.28, 2.55] straddling zero widely, PSR 0.606, effective_daily_obs 25

Cost-heavy and capacity-dead: commission 24.3% of gross + impact 15.7% (~40% of gross consumed), capacity_usd only ~$4.05M; exposure_pct 2.24% means capital is almost entirely idle

Same-asset BTC HL-vs-Binance perp basis is normally ~1-2 bps, so a >15bps 'discount' tail risks being stale-print / bar-misalignment artifact as much as a real dislocation (L57/L44 cross-venue basis family)