Drift Detection

A model trained on 2018-2022 data may silently fail when the 2023 distribution shifts. Systematic drift detection catches degradation before it becomes a drawdown.

The Problem

Financial distributions are non-stationary. Volatility regimes change, correlations spike during crises, and feature distributions shift as markets evolve. A model deployed without monitoring can trade on stale assumptions for months. The three drift types - data drift (input distributions change), concept drift (feature-target relationship changes), and prior drift (target distribution changes) - each require different detection strategies. By the time performance metrics visibly decay, the damage is already done.

The Pattern

WRONG

# Deploy model and check performance monthly - too late
model = train_model(X_train, y_train)
# ... 3 months later ...
print(f"Live Sharpe: {live_sharpe:.2f}")  # Already lost money

CORRECT