Formula
y = β₀ + β₁x₁ + … + βₙxₙ + ε
Why Start with a Linear Model
- Simple to understand and communicate.
- Provides a transparent baseline for comparison with more complex models.
Data and Features
- Explore distributions, correlations, multicollinearity and outliers.
- Encode categorical variables and use domain knowledge for feature engineering.
- Use correlation analysis and VIF to identify redundant features.
- Confirm that production features are available, stable and appropriate.
Training and Evaluation
- Select an appropriate loss function and regularisation method.
- Evaluate MSE, RMSE, MAE and R-squared.
- Use cross-validation, residual analysis and learning curves.
Interpretation
- Review coefficients, scatter plots, partial dependence and SHAP values.
- Do not interpret predictive relationships as causal evidence.