Does Algorithmic Directional Trading in Cryptocurrencies Survive Transaction Costs? An Empirical Study with Public Binance Futures Data

Abstract This study empirically evaluates whether algorithmic directional trading in cryptocurrencies possesses an exploitable edge after realistic transaction costs. We evaluated 60 configurations, approximately 45 of which were directional, organized into eight methodological categories: supervised machine learning, genetic algorithms, microstructure, multi-temporal ensembles, latent state models, asymmetric beta, quantized waves, and horizon analysis. Exclusively public data from Binance Futures was utilized. The protocol includes walk-forward validation with horizon-scaled purging, a point-in-time (PIT) universe including delisted assets, and round-trip (RT) costs of 0.20%. Five robustness criteria were pre-registered before observing any results. The ceiling for directional AUC using OHLCV features is ~0.53. Microstructure models reached an AUC of 0.540. Horizon analysis shows that the 90-day AUC reaches 0.710 under the legacy protocol, but drops to 0.644 under correct purging, and collapses to 0.528 under non-overlapping subsampling—an unbiased estimator indistinguishable from random chance. The operational backtest yields a Calmar ratio of 0.115. Block bootstrap confirms that no horizon produces an AUC significantly different from chance. The Deflated Sharpe Ratio (DSR), applied to the best candidate (observed Sharpe of 0.237, N=45 directional trials), results in 0.0000. As a methodological control, a canonical momentum rule achieves a Calmar of 1.101, but fails the pre-set institutional criteria. None of the complex systems even reach that baseline level. We conclude that, using public data and retail costs, there is no robust directional edge that meets institutional standards, and that complex algorithmic systems fail to outperform simple rules found in the literature.