AI Audit Finds Crypto Trading Models Unprofitable on Binance Spot

An extensive AI-assisted audit has concluded that candle-based machine learning models designed to predict cryptocurrency price extrema or short-horizon outcomes on Binance Spot are consistently unprofitable when real-world costs are factored in. Despite some models showing decent predictive performance in terms of ROC AUC, their average precision was low, leading to significant losses in simulated trading. The audit, which involved human-supervised AI agents for literature retrieval and critique but not trading decisions, tested various strategies over multiple cycles. The strongest performing models still resulted in substantial negative returns, with one selector losing 6.72% over 19 cycles and others underperforming a simple buy-and-hold strategy. The findings suggest that event-ranking performance alone does not translate into positive executable policy value in this highly volatile and cost-sensitive market.