When a P&L attribution or document-risk model "doesn't train," the failure is usually readable from the train and validation traces long before you rewrite the architecture: underfit, overfit, exploding steps, data leakage, or a broken metric. This note turns those curves into a diagnosis checklist and a hyperparameter-search protocol so each experiment answers one question under a frozen split.
In this lesson
- Learning objectives
- What to log every run
- Curve patterns and first repairs
- Error analysis before more layers
- Hyperparameter search without burning the test set
- Failed-training quick checklist
Before you start
- Normalization, Residual Connections, and Stable Deep Architectures
- The Supervised Learning Problem and the ML Pipeline
The full lesson, with worked examples and exercise solutions, is for members.
Read the full lessonNew to SelfMadeQuant.com? See membership options or explore the curriculum.