Predictive Extrema, Unprofitable Policies: An AI-Assisted Audit of Candle-Based Binance Spot Timing Models
Why AI models that predict crypto price swings still lose money in practice
An audit of machine-learning models designed to predict extreme price movements and time trades on Binance found that none of them beat simply holding cryptocurrency—most lost 1–44% over their test periods even though the models accurately identified price patterns. The best-performing model predicted local price extrema with high precision but generated trading strategies that underperformed buy-and-hold by 2.80 percentage points after accounting for trading costs.
This work exposes a common gap in cryptocurrency and algorithmic trading: accurate price prediction doesn't automatically translate into profitable trading decisions. The findings suggest that even when AI systems correctly identify when prices will peak or bottom, the costs and timing of acting on those predictions make the trades net negative—a crucial reality check for traders and firms considering expensive machine-learning systems to automate their market timing.