Simple, naive strategies that any ML model must beat convincingly to prove it has real predictive power. If your complex XGBoost model can't beat "always predict up," it's not worth the complexity.