Probability-based portfolio in crypto-perpetual futures market

Machine learning models often struggle to outperform traditional approaches in out-of-sample return prediction because financial returns exhibit a low signal-to-noise ratio. We argue that achieving strong portfolio performance does not require highly accurate point forecasts or ever more complex models; instead, it depends on how predictive signals are translated into tradable positions. We propose ProSP, a model-agnostic probability-based long-short strategy that ranks cryptocurrency perpetual futures by predicted tail probabilities. ProSP dual-sorts assets into a long leg with high top-tail probability and a short leg with high bottom-tail probability, and removes overlapping signals, which arise endogenously under elevated volatility, to preserve directional clarity. We provide theoretical conditions under which tail probabilities are valid ranking statistics and characterize the structural origin of probability intersection. Empirically, we focus on perpetual futures, a setting with particularly low signal-to-noise. In a rolling out-of-sample test from January 2021 to May 2025, ProSP consistently outperforms regression and hard-label classification benchmarks across feature sets, horizons, and model classes (Decision Trees, Random Forests, XGBoost, and neural networks), delivering substantially higher Sharpe ratios and market-neutral alpha. The results show that even modestly predictive machine learning models can produce economically meaningful performance when embedded in a probability-based portfolio construction framework.