Monitoring a Live Trading Model

A model trained in January fails in May if the regime shifted. Monitor: feature means (drift alert if >2 std), out-of-sample Sharpe rolling, prediction calibration (do P=0.7 events happen 70% of time?). When drift detected, retrain on recent window. Silent decay is the retail killer -- the model looks fine while bleeding. Publish a monitoring dashboard concept; it separates a system from a lucky notebook. A model you cannot monitor is a liability wearing a backtest.