Strategies

BTC Spot-Perp Premium Catch-Up Long (Cross-Venue Basis Signal)

Outcome: Abandoned

BtcSpotPerpPremiumCatchUpLongCrossVenue

Outcome Summary

BtcSpotPerpPremiumCatchUpLongCrossVenue traded BTC spot long off a perp-premium signal, betting that leveraged perp positioning leads spot with a 1-3 day catch-up lag, filling the under-quota cross-venue dimension without funding exposure. The idea was well-coded — 460 clean long-only entries, timestamp-aligned legs, active trend filter — but the edge was negative in every year (PF 0.86, Sharpe -1.43, expectancy -$134, -23.3% return) with fees eating 19.2% of gross. The analyst ruled it a refuted premise rather than a tuning problem: the basis converges via the perp falling, not spot rising, so buying spot on premium catches the losing leg, and recommended fading the premium instead. It ended after one iteration as abandoned, never advancing to optimization or risk review.

Strategy report

Hypothesis

A long-only, cross-venue strategy that trades BTCUSDT BINANCE_SPOT (CASH long-only, ~0.20% RT fees) using the BTCUSDT.BINANCE perpetual futures price as a leading microstructure signal. The mechanism: when perp price trades at a meaningful PREMIUM to spot (basis_bps >= threshold), this indicates leveraged long positioning in the perp market that is bullishly information-carrying — spot tends to follow with a 1-3 day lag as the basis normalizes. Buying spot on perp premium captures the catch-up move without shorting (so it works on a CASH spot account) and without holding the leveraged perp (so it dodges funding-rate exposure). Critically distinct from every existing portfolio entry: (a) NOT a classical indicator (MACD/OBV/SAR/HA/hammer all empirically refuted); (b) NOT a single-instrument price-action pattern; (c) NOT a volume-z, skewness, or range-geometry mechanism (those are in pipeline); (d) IS a cross-venue microstructure signal — directly addressing the most under-quota dimension (cross_venue 2.8% vs 15% target). Both required instruments are confirmed working in current pipeline: BtcFourHourVolumeBreakoutLong validates BTCUSDT.BINANCE perp 4H bars, BtcSpotDrawdownAccumulationLong validates BTCUSDT.BINANCE_SPOT bars. OHLCV-only: no funding/OI/liquidations supplementary-data dependency. Trend-filtered to avoid the Keltner-style 'fade the downtrend' failure.

Implements the hypothesis directly: a cross-venue microstructure edge (perp premium leads spot) expressed as a long-only spot trade. Venue routing is deliberate -- BINANCE_SPOT for the traded leg because the strategy is long-only, uses no leverage, and the edge is NOT funding-based (so spot's CASH account and ~0.20% RT fees are correct, not a futures venue); the BINANCE perp is included only as a read-only signal source (extra_instrument), never traded, so there is zero funding/leverage exposure. Primary = spot so get_account_equity reads the spot CASH account and the base _submit_entry/_submit_exit operate on the long-only leg (BUY to enter, SELL to close -- never shorts). The uptrend SMA filter avoids the documented 'fade the downtrend' failure mode. Fee-aware: spot RT ~0.20%; take-profit (1.5%) and stop (1.0%) both dwarf fees, and basis-normalization exit captures the catch-up cleanly. Signal (basis_bps) is returned continuously so it varies bar-to-bar (Layer-2 safe); the threshold + trend gate are in should_enter. Default entry of 5 bps is set to reliably generate the convergence trade; the optimizer can widen it for a higher-conviction premium. leverage=1.0 (spot caps at 1x; not referenced as >1, so no leverage-unused gate).

No edge in any regime — not an optimization candidate. The strategy loses money every single year (2020 -2.6%, 2021 -21.7%, 2023 -7.4%, 2024 -0.7%, 2025 -1.6%) with profit_factor 0.86, Sharpe -1.43, Sortino -2.50, negative alpha -0.036, and negative per-trade expectancy (-$134/trade, avg_trade_return well below the >0.25% spot fee bar; commissions were 19.2% of gross). The mechanism is implemented correctly (460 long-only spot entries matching the hypothesis, clean cross-venue alignment, signals firing 460/460), so this is a PREMISE failure, not a code bug: the 'perp premium leads spot's 1-3 day catch-up' thesis is empirically refuted — basis convergence occurs largely via the perp reverting down to spot, not spot rising to the perp, so buying spot on premium systematically catches the losing half of the convergence. Optimization tunes parameters toward a robust positive region; with a uniformly negative result across all regimes there is no such region — tuning would only curve-fit noise to a single window and waste the 2-hour compute budget. The premise (premium=bullish-leading) is the dead component; if the cross-venue basis dimension is still wanted, the Research Lead should reframe the mechanism (e.g. fade perp premium / mean-revert the basis, or test premium as a SHORT-spot-avoidance filter) rather than re-running the same long-the-premium logic through the optimizer.

Implementation

Long-only cross-venue basis strategy. Trades BTCUSDT on BINANCE_SPOT (CASH) using the BTCUSDT.BINANCE perpetual as a leading microstructure signal. Computes basis_bps = (perp_close - spot_close)/spot_close * 10000 each 4H bar (the perp leg is timestamp-aligned to spot by the base class's same-timeframe sync barrier, so the basis is never computed on a stale leg). When the perp trades at a meaningful PREMIUM (basis_bps >= entry threshold) AND spot is in an uptrend (close > SMA(trend_period)), it buys spot to capture the ~1-3 day catch-up move as the basis normalizes -- no shorting (CASH-legal) and no perp position (no funding exposure). Exits on take-profit, stop-loss (capital-relative %), basis normalization (basis_bps <= exit threshold), or a max-hold cap. Sizing is capital-relative and CASH-capped (notional <= equity * position_frac, no leverage). OHLCV-only, no supplementary data.

Backtest Review

Sharpe: -1.43
Total return: -30.93%
Max drawdown: 35.38%
Trades: 460
Win rate: 45.2%
Profit factor: 0.86

Mechanism is implemented correctly and matches the hypothesis: 460 long-only spot entries, 0 shorts, signals fire cleanly (460/460 submitted, 0 dropped), cross-venue spot/perp legs are timestamp-aligned, trend filter active.

Ample sample size (460 trades over ~6 years) and good cross-venue dimension coverage — the idea was well-formed and well-coded.

Negative edge in EVERY regime: annual returns negative for 2020, 2021, 2023, 2024, 2025 — not a window-specific or cliff problem that tuning can fix.

Profit factor 0.86 (<1.0), Sharpe -1.43, Sortino -2.50, alpha -0.036, expectancy -$134/trade — the signal has no positive edge to optimize toward.

Fee-fragile: commissions were 19.2% of gross on a spot CASH account (~0.20% RT); even a marginal raw edge would be consumed. avg_trade_return is negative, far below the >0.25% spot bar.

Premise appears empirically refuted: perp premium does not lead spot's catch-up on a 1-3 day horizon — basis convergence happens largely via the perp reverting down, so buying spot on premium captures the wrong leg.

negative in all 5 years

positive OOS

Outcome Summary

Perp premium does not lead a bullish spot catch-up: cross-venue basis convergence occurs mainly through the perp reverting down rather than spot rising, so longing spot on premium systematically captures the wrong half of the convergence — if the cross-venue basis dimension is still wanted, reframe to fade the perp premium / mean-revert the basis rather than re-run the long-the-premium logic.

It was abandoned at the pre-optimization backtest-review gate (verdict: abandon) as a premise failure, not a code bug — the 'perp premium leads spot's catch-up' thesis is empirically refuted (basis convergence happens largely via the perp reverting down to spot, so buying spot on premium catches the losing leg) — and with a uniformly negative result across all regimes there is no robust region to optimize, so optimization and all later stages were never reached.

A long-only cross-venue basis microstructure strategy — trading BTCUSDT on BINANCE_SPOT (CASH, long-only) while reading the BTCUSDT.BINANCE perpetual as a leading signal: when the perp traded at a premium to spot (basis_bps ≥ threshold) within an SMA uptrend, buying spot to capture the expected 1-3 day catch-up as the basis normalizes, exiting on take-profit, stop-loss, basis normalization, or a max-hold, with no funding exposure and OHLCV-only data.

Over ~6 years and 460 long-only spot entries (signals firing cleanly, 0 dropped) it had a negative edge in every regime: profit factor 0.86, Sharpe -1.43 (CI [-3.26, 0.32]), Sortino -2.50, expectancy -$134/trade (avg_trade_return -0.13%, below the >0.25% spot bar), total return -23.3%, 33.7% max drawdown, negative annual returns in 2020/2021/2023/2024/2025, with commissions at 19.2% of gross on the ~0.20% RT spot account.

Backtest and paper results are hypothetical. Trading involves risk of loss.