Tree depth and sample fragmentation

Why can a deep regression tree fit weak financial signals extremely well in training but poorly in later periods?

Repeated splitting creates small leaves whose averages can fit sample noise and unusual observations. Future samples may not share those patterns. Constrain depth and leaf sizes, compare against simple baselines, and tune with chronological validation. Inspect the effective sample size within leaves, not just total row count. Correlated stock-date observations can make a numerically large leaf contain little independent information.