Derivatives data is one of the most underutilized edges in crypto trading. While retail traders watch price charts, the open interest, funding rates, long/short ratios, and spot-futures basis tell a different story: the story of leveraged positioning, crowding, and the conditions that precede liquidation cascades. We set out to build strategies that trade on this data. Four strategies entered the pipeline. One survived.
The Four Candidates
We built four distinct strategies that use derivatives data from Binance perpetual futures and Coinglass aggregation.
The first was oi_momentum: a strategy that combines open interest direction with price direction. Rising OI with rising price suggests new longs entering the market. Rising OI with falling price suggests new shorts. The strategy attempts to ride the momentum of leveraged positioning.
The second was funding_contrarian: a strategy that fades extreme funding rates. When funding is highly positive (longs paying shorts), the market is crowded long. When funding is highly negative, the market is crowded short. The strategy takes the opposite side of the crowd at funding extremes.
The third was leverage_composite: a multi-signal strategy that combines open interest change, funding rate extremes, and long/short account ratios into a composite score. A trade is generated when the composite score reaches a threshold, requiring agreement across multiple indicators.
The fourth was basis_convergence: a strategy that trades the spot-futures basis. When the basis is abnormally wide (futures trading at a premium to spot), the strategy shorts futures expecting convergence. When the basis is abnormally narrow or negative, it goes long.
Tournament: First Filter
The tournament screening tested all four strategies at default parameters across symbols with derivatives data. Several important symbols (BTC, ETH, XRP, BNB, ADA) had no OI/LSR/basis data in our database and produced zero trades, which was expected.
The tournament results were modest but not discouraging. oi_momentum showed Sharpe between 0.5 and 1.5 on a handful of symbols. funding_contrarian was scattered, with some symbols positive and some negative. leverage_composite showed the most consistent results, with 4 of 7 testable symbols above Sharpe 1.0. basis_convergence was marginal at best.
Leverage_composite's consistency was notable. While the absolute numbers were not impressive, the fact that it was positive on more symbols than any other candidate suggested a broader edge rather than symbol-specific fitting.
Sweep: Parameter Calibration
The parameter sweep was extensive. For leverage_composite, we tested grids across oi_lookback (7 to 28 days), oi_threshold (0.01 to 0.05), funding_extreme (0.0001 to 0.0005), lsr_crowded_long (0.55 to 0.75), and min_score (1 to 3). That is hundreds of combinations per symbol, run across the workers for speed.
The sweep revealed clear parameter preferences. The best configurations clustered around oi_lookback=14, oi_threshold=0.024, funding_extreme=0.00036, lsr_crowded_long=0.65, min_score=2. This clustering is important: when optimal parameters form a tight neighborhood, the edge is more likely to be real than when the best configuration is an isolated point.
For oi_momentum, the sweep produced higher peak Sharpe ratios (3.34 on INJ) but with no clustering. The best parameters on INJ were completely different from the best on ARB, which were different from WIF. This is a sign that the sweep is finding symbol-specific patterns rather than a general edge.
Funding_contrarian showed a dead end. The best Sharpe was 1.94 on SHIB, but the strategy required extreme parameter values to produce signals, and even then, trade frequency was too low for statistical confidence. We classified it as not deployable.
Basis_convergence showed partial promise on APT but failed to produce consistent results across symbols. The spot-futures basis is noisy and the mean-reversion dynamics are slower than we hoped.
Phase 2: Targeted Refinement
For leverage_composite, Phase 2 generated a plus/minus 20% grid around the sweep winner: oi_lookback from 11 to 17, oi_threshold from 0.019 to 0.029, and so on. This higher-resolution search confirmed the parameter neighborhood and refined the winning configuration slightly.
The Phase 2 winner: oi_lookback=14, oi_threshold=0.024, funding_extreme=0.00036, lsr_crowded_long=0.65, min_score=2. Almost identical to the sweep winner, which is a good sign. It means the sweep found the right neighborhood on the first pass.
We ran Phase 2 on the workers, as we always do for compute-intensive jobs. The job was registered on the coordinator (Mac), then local processes were immediately killed and the workers handled all execution.
Validation: The Decisive Test
Derivatives data has a shorter history than price data. Our Coinglass Hobbyist plan provides approximately 6 months at 4-hour granularity. This means we cannot use our standard 5-year validation periods. Instead, we defined three sub-periods within the available data: October-December 2025, January-February 2026, and March-April 2026.
The validation results were decisive.
leverage_composite on ARB/USDT: ROBUST. Profitable in all 3 periods. Average Sharpe 2.91. The strategy captured leveraged positioning shifts consistently across a volatile correction, a recovery, and a consolidation phase.
leverage_composite on OP/USDT: ROBUST. Profitable in all 3 periods. Average Sharpe 1.89. Lower absolute performance than ARB but consistent across regimes.
leverage_composite on WIF/USDT: strong but not perfect. Profitable in 2 of 3 periods. Average Sharpe 3.02. The one negative period was marginal (Sharpe -0.15), suggesting bad luck rather than strategy failure.
oi_momentum: failed validation. Only INJ showed edge in 2 of 3 periods. The sweep Sharpe of 3.34 was regime-fitted, exactly as we expected from the lack of parameter clustering.
basis_convergence: failed validation. Only APT showed partial results (2 of 3 periods). The strategy was negative in 2 of 3 periods overall.
funding_contrarian: we ran validation for completeness despite the dead-end classification from the sweep. BTC produced Sharpe -4.26. Severe overfitting confirmed.
Why Composite Beats Single Signal
The key insight from this process is that composite signals are more robust than single-factor signals. oi_momentum uses one signal (OI direction relative to price direction). funding_contrarian uses one signal (extreme funding rates). Both are legitimate edges in theory, but both are noisy enough that they generate false signals frequently.
leverage_composite combines three signals: OI change, funding rates, and long/short ratios. A trade requires min_score=2, meaning at least two of the three signals must agree. This consensus mechanism filters out the noise that plagues single-factor strategies.
When OI is rising while funding is highly positive and the long/short ratio shows crowding, you have triple confirmation of a directional imbalance. The probability that all three indicators are simultaneously giving false signals is much lower than the probability that any one of them is wrong.
This is the same principle that makes our multi-indicator consensus strategy work in the technical analysis category. Agreement across uncorrelated signals is a stronger foundation for edge than any single signal, no matter how theoretically compelling.
Deployment
Based on the validation results, we deployed leverage_composite to paper trading on April 2, 2026. Three bots were created: ARB/USDT, OP/USDT, and WIF/USDT, each with $1K allocation, running on 1-hour timeframes.
The parameter configuration differs slightly by symbol. ARB and WIF use the standard Phase 2 winners (oi_lookback=14, oi_threshold=0.024, funding_extreme=0.00036, lsr_crowded_long=0.65). OP uses a slightly higher oi_threshold of 0.03, reflecting its different OI dynamics.
These three bots bring our total to 45, deployed across 9 strategies. leverage_composite is our first derivatives strategy in production, opening a new category of alpha that is uncorrelated with our price-based strategies. The derivatives signals respond to leveraged positioning changes, not price pattern recognition, so they diversify the portfolio's return drivers.
The full pipeline from initial strategy construction to paper trading deployment took approximately two weeks. Four strategies built, four tournaments run, four sweeps completed, one Phase 2 refinement, one validation, three bots deployed. The ratio of strategies tested to strategies deployed (4:1) is typical for our process and reflects the reality that most trading ideas do not survive contact with data.