Early stopping consumes validation data

You stop boosting when validation error is lowest and report that minimum as an unbiased performance estimate. What is the issue?

The validation data selected the stopping iteration, so the minimum is optimistically selected. Use that window for stopping but evaluate the resulting training procedure on an independent outer period. Repeat the stopping rule within each historical fold, with preprocessing and any label purging fitted appropriately. Early stopping is a hyperparameter-selection mechanism even if its implementation is automatic.