XGBoost retraining, fine-tuning debate tackles clickstream drift

This is really a production-ML question disguised as a training-method question. For XGBoost, rolling retrains are usually the safer default; for the bandit layer, incremental updates still make sense.

• –XGBoost does support training continuation, but daily tree-stacking can make versioning and drift diagnosis messy.
• –The 30/90/180-day weighting already bakes in recency; the real work is backtesting window sizes against recent holdouts.
• –Retrain on schedule or when drift and performance metrics slip, not just because new data landed.
• –Keep Thompson sampling or LinUCB incremental, since those methods are built to absorb feedback online.
• –Transfer learning is mostly the wrong mental model here; this is about retrain cadence and monitoring, not model reuse in the neural-net sense.