Convergence analysis checks whether an iterative training procedure approaches stable loss and improves on established baselines. In forecasting, it becomes relevant when experimenting with gradient boosting, not with simpler baseline models. section 9.4.1
A practical convergence check asks:
- Does the model converge as iterations increase? section 9.4.1
- Does it beat baseline quantile loss, mean absolute percentage error, and related metrics? section 9.4.1
- Does the curve reveal underfitting or overfitting? section 9.4.1
- After convergence is established, select a rough tree-count grid such as 500, 1,000, 2,000, 3,000, and 5,000 trees. section 9.4.1
Retail demand forecasting
A forecasting system uses six quantile losses and corresponding models for the 1.5th, 25th, 50th, 75th, 95th, and 99th quantiles. Compare each gradient-boosting model with constant baselines, inspect loss across iterations, and fix a sufficient number of trees for later experiments. section 9.4.1
