BtcCrossVenuePerpBasisMeanReversionHlBinance4H

BTC Inter-Exchange Perp Basis Mean-Reversion (Hyperliquid vs Binance, Cross-Venue Market-Neutral, 4H Bars)

Cross-venue market-neutral basis arbitrage between BTCUSDT perpetual on BINANCE (USD-M) and BTC perp on HYPERLIQUID. The strategy continuously monitors the price ratio between the two venues' BTC perpetuals on 4-hour bars and trades the convergence of transient mispricings. When the basis spread (HL_price / Binance_price - 1) reaches an extreme threshold (>= +0.45% or <= -0.45%), enter a delta-neutral pair: LONG the cheaper venue's perp + SHORT the more expensive venue's perp in equal USD notional. Position is held until the basis mean-reverts to a tight band (|basis| < 0.10%) at which point both legs are closed simultaneously. The strategy is fundamentally distinct from funding-carry / funding-reversion strategies already in the pipeline — it captures PRICE-LEVEL divergence between two perp markets, not funding rate cash flows. Both venues are highly liquid for BTC and the strategy is independent of BTC's directional movement. Uses 4-HOUR bars on both venues (NOT 8-HOUR which has had data refill issues on BINANCE in recent failures). Fills 4 under-represented portfolio buckets simultaneously: cross_venue (4.3% vs >=15%), HYPERLIQUID (4.2% vs >=20%), pairs (13.5% vs >=15%), and long_short (9.1% vs <=55% long_only). The strategy directly addresses the most acute portfolio diversification gaps.

Iteration 2 fix targeting the Layer-3 'trades clustered in a single May-2026 week' failure while preserving everything that already passed Layers 1-2 (imports, class structure, signal scaffolding, neutral-pair logic all unchanged). Verified the data reality: the HL BTCUSD 4-HOUR catalog file spans 2024-01-22 to 2026-05-22 and Binance BTCUSDT 4-HOUR spans 2019-12-31 to 2026-06-25, so there IS a genuine ~2.3-year contemporaneous overlap at 4H — the prior single-week clustering was NOT a coverage gap but a threshold-scale problem. A 0.45% de-meaned divergence between two liquid BTC perps is large and only occurred during the turbulent May-2026 window, hence the cluster. Recalibration: (1) entry_threshold_pct 0.45 -> 0.15 (the realistic BTC perp-perp divergence band per the feedback's 0.10-0.25% guidance); (2) basis_mean_window 120 -> 36 bars so the rolling mean tracks the current regime and the transient component is isolated/threshold-crossed far more often across the whole overlap; (3) exit_threshold_pct 0.10 -> 0.04 to take convergence profit before round-trip fees erode it; (4) max_hold_periods 0 -> 30 to guarantee turnover and bound basis risk. These changes only touch parameter defaults/config values, so Layers 1-2 stay green, and they spread dozens of entries across the genuine 2024-2026 overlap to support optimization.

Not worth 2 hours of optimization. This is already iteration 2 (thresholds lowered to chase frequency) yet the BTC HL-vs-Binance basis produced only 10 entries / 20 trades over the entire overlap, concentrated in 2026-05 (2025 was negative), with exposure 1.63%. The two highly-liquid BTC perps are too tightly coupled for a tradeable post-cost edge: profit_factor 1.06, per-trade edge ~0.10%, and market impact already consuming 40.6% of gross PnL at a tiny $608k capacity. This is the structural 'two liquid perps barely diverge -> edge below costs' problem — lowering the threshold further only adds impact/fee drag, so parameter optimization cannot fix it, and 10 concentrated trades would just fit noise. The cross-venue basis-MR mechanism itself is sound and could be reframed onto a more divergent / less-liquid cross-venue pair (a Research Lead hypothesis change), but the BTC HL-Binance target as specified has no deployable edge.

Delta-neutral cross-venue BTC perp basis mean-reversion: monitors the de-meaned price basis (HL_price/BINANCE_price - 1) between BTCUSD.HYPERLIQUID and BTCUSDT.BINANCE on 4H bars. When the de-meaned spread exceeds 0.15% it opens a market-neutral pair (long the cheaper venue, short the richer, equal USD notional) and closes both legs when the spread reverts inside 0.04% or after a 30-bar max hold. Captures transient price-level divergence between two liquid BTC perps, independent of BTC direction; uses no supplementary data.

Strategy is mechanically functioning: produces balanced delta-neutral pairs (10 long / 10 short legs), matching the hypothesized cross-venue basis mechanism

Market-neutral construction confirmed (beta 0.0025, benchmark_correlation 0.11); benchmark_meaningful=false correctly flags buy-hold as inapplicable

Basis entry magnitude (de-meaned 0.15%) is plausible, not a >0.5% stale-leg/misalignment artifact

Only 10 entries / 20 trades over the full ~2024-2026 HL-Binance overlap despite iteration 2 already lowering thresholds to chase frequency — far too few to optimize meaningfully (225 trials would fit noise)

Edge is concentrated in a single window (2026-05); 2025 is NEGATIVE (-1.8%) and exposure is just 1.63% of the time

Marginal post-cost edge: profit_factor 1.06 (<1.2 bar), avg_win $932 ~ avg_loss $880, per-trade edge ~0.10% — fees consume nearly all gross

Market impact eats 40.6% of gross PnL with capacity only ~$608k — edge real only at toy scale

Sharpe CI straddles 0 widely (sharpe_ci_low -7.54, high 10.86); the 1.58 Sharpe is not distinguishable from noise on this sample

Do NOT proceed to optimization — only 3 entries exist, all in a single week (May 2026), which cannot support parameter fitting. Fix two things first: (1) DATA COVERAGE — confirm the HYPERLIQUID BTCUSD 4H bars actually load and overlap the BINANCE BTCUSDT series across the full backtest range; entries firing only in May-2026 strongly suggests _extra_bars is empty/unaligned for most of the period (or the alignment barrier rejects nearly all bars). Print/verify the count of bars where both legs are contemporaneous; if HL history only starts recently, restrict the backtest window to the genuine overlap so the sample is real rather than 99% dead bars. (2) THRESHOLD CALIBRATION — once coverage is confirmed, lower entry_threshold_pct to the realistic BTC perp-perp divergence scale (try 0.10-0.25%) and/or shorten basis_mean_window so the signal fires often enough (target dozens+ of trades) to be optimizable. Also sanity-check the May-2026 entry spreads (-0.67%/+0.46%/+0.69%): a >0.45% divergence between two liquid BTC perps is large — verify these are real contemporaneous quotes and not a stale HL leg. Re-backtest and only return for optimization once there is a statistically meaningful trade count over a genuine data-overlap window.

This strategy proposed a market-neutral cross-venue arbitrage trading the convergence of transient price dislocations between the BTC perpetual on HYPERLIQUID and BINANCE on 4H bars, and by iteration 2 it had already lowered its entry threshold from 0.45% to 0.15% to chase trade frequency. The backtest confirmed the mechanism worked — balanced 10 long / 10 short delta-neutral pairs with beta ~0.0025 — but it generated only 20 trades over the multi-year overlap, with the edge concentrated in a single 2026-05 window and 2025 actually negative. With profit factor 1.06, exposure 1.63%, and market impact eating 40.6% of gross PnL at just ~$608k capacity, it failed every backtest-review gate. The analyst returned an 'abandon' verdict before any optimization, concluding the two liquid BTC perps are too tightly coupled for a deployable post-cost edge, though the basis-MR mechanism itself could be reframed onto a more divergent cross-venue pair.

Two highly liquid BTC perpetuals are too tightly coupled to diverge enough for a post-cost edge — lowering thresholds only adds fee and impact drag, so this mechanism needs a more divergent or less-liquid cross-venue pair to be tradeable.

It was abandoned at the pre-optimization backtest-review gate, where the analyst issued an 'abandon' verdict after it failed every gate (total_trades 20<30, profit_factor 1.06<1.2, exposure 1.63%<5%, impact_cost 40.6%>25%); analyst, risk, and optimization stages were never reached.

A delta-neutral cross-venue basis mean-reversion on BTC perpetuals, going long the cheaper and short the richer of the HYPERLIQUID and BINANCE BTC perps on 4H bars whenever the de-meaned price spread reached an extreme, to capture transient inter-exchange dislocations independent of BTC direction.

Over the backtest it produced only 10 entries / 20 trades with a 50% win rate, total return 2.28%, Sharpe 1.58 (CI -7.54 to 10.86), profit factor 1.06, and exposure of just 1.63% of the time, with edge concentrated in a single 2026-05 window (2025 was negative at -1.8%). Market impact consumed 40.6% of gross PnL at a capacity of only ~$608k.