🎩 Edward Thorp, inventor of card counting and pioneer of quantitative trading, developed an $800M market-neutral pairs-trading strategy to profit from relative price movements between correlated assets (e.g., ETH/ETC). This statistical arbitrage approach involves simultaneous long and short positions, capitalizing on momentary divergences independent of market direction. The video details rebuilding and enhancing this strategy for 2025 using the Jesse framework.
The foundational setup used Jesse within a Jupyter notebook, fetching 50-minute Binance perpetual futures data for simultaneous long/short entries. Prices were converted to returns for normalization. The strategy's core relied on the Z-score of the spread (asset return difference), generating signals upon crossing predefined thresholds (e.g., $\pm 1.2$ from the mean) and closing upon mean reversion. Asset co-integration was also tested.
Implemented in Jesse, the Z-score was calculated from the last 200 candles' spread. Position sizing utilized util.calculate_alpha_beta. Initial simulated results for BTC/SOL without fees looked fantastic (e.g., -5% max drawdown); however, integrating realistic trading fees (Binance taker fees plus slippage) led to a devastating -95% capital loss, underscoring the critical need for robust filtering.
Key improvements were introduced to achieve viability:
- Bollinger Bands (BB): Replaced static Z-score thresholds. BBs were applied to the spread; entries were triggered when the spread moved beyond the Upper Band (short) or below the Lower Band (long), and exits when it crossed the Middle Band. This offered dynamic adaptation to volatility.
- ADX Indicator: An entry filter ensuring trades occurred only when market volatility (ADX) was above 30, preventing entries during flat, low-opportunity periods.
- Trend Filter (KAMA): Filtered entries based on the primary asset's trend; long positions when price was above KAMA (uptrend), short positions when price was below KAMA (downtrend), aligning trade bias.
- P&L Filters (Most Important): These addressed "unworthy" trades and managed positions dynamically.
is_worth_opening: An entry filter requiring the absolute spread value to be at least 1.5%. This ensured potential profits covered transaction costs, noting adjustment for asset volatility (e.g., 4% for meme coins).is_worth_closing: A dynamic profit-taking rule, closing positions when the total P&L (both legs) reached a 1.5% profit.time_to_stop: A comprehensive stop-loss, liquidating the entire pair position when the collective P&L reached a -5% loss, crucial for pair strategies where individual asset stop-losses are unsuitable.
These improvements dramatically transformed the strategy. ✅ The enhanced model yielded a +18% return over six months in 2025 (BTC/SOL), with a maximum drawdown of only -2% and an impressive 87% win rate, all after accounting for fees and slippage. This represented a substantial shift from the initial losses.
Final Takeaway: Despite the positive turnaround, important caveats remain. The backtested 2025 period was acknowledged as "cherry-picked," with 2024 performance being less favorable. Real-world slippage can exceed simulations, particularly with volatile assets or lower liquidity, necessitating higher fee allowances. For smoother equity curves and improved metrics (e.g., lower max drawdown), diversifying across multiple pairs (4, 6, 8, or 10+) simultaneously is strongly recommended. These refinements, alongside continuous development and strategic diversification, are key to building robust statistical arbitrage strategies.
