Error analysis turns model mistakes into a guide for the next system update. It examines learning behavior during training and residual distributions after prediction, revealing weak spots, trends, and recurring patterns that metrics alone can hide. section 9.0 section 9.5
Use a practical sequence:
- Compare candidate models with constant or dummy baselines before interpreting improvements. section 9.4.1
- Inspect learning curves for convergence, baseline improvement, underfitting, and overfitting; stop deeper analysis when training behavior is unsound. section 9.4.1 section 9.5
- Relate error size to model complexity, feature count, lag choices, or window sizes to choose a useful capacity. section 9.4.1
- Inspect high- and low-loss samples, retain metadata, form hypotheses about each error, and group recurring causes for follow-up. section 9.4.2
Two operational uses
For Supermegaretail, compare linear and gradient-boosting forecasts with sales-history baselines, then tune trees, lags, and windows using loss-versus-complexity plots. For PhotoStock, inspect high-loss search results alongside query, image, language, geography, category, and tag metadata; group causes such as blurry images, overoptimized descriptions, or overly short queries. section 9.4.1 section 9.4.2
Treat error analysis as a design requirement, not a final audit. Reapply it whenever a new improvement is planned, because changing data, features, losses, or model capacity can change the error pattern. section 9.4 section 9.4.2
