How to teach a crypto scanner not to lie to itself: The ML part of AltScanner

The developers of AltScanner have introduced an approach to integrating machine learning into a trading scanner, focusing on data quality and preventing model 'self-deception.' Instead of training algorithms directly on market candles, the team started by building a robust system for capturing market states and a delayed labeling mechanism. This avoids common pitfalls like data leakage or incorrect interpretation of open candles. The ML circuit operates alongside the primary two-axis evaluation system (D for direction, A for activity), analyzing 13 tokens on the USDT perpetual market across 15, 60, and 240-minute timeframes. The model does not replace existing indicators but complements them by studying correlations between features and price movements. This approach ensures system interpretability and improves prediction accuracy while maintaining the transparency of trading signals for platform users.