underfitting

Underfitting

When a model is too simple to capture the true patterns in the data — produces high error on both training and test data.

When you'd see it: A linear model applied to a genuinely non-linear problem. A shallow decision tree on complex data. Any situation where adding more training data doesn't improve performance — the model's capacity is the bottleneck, not the data.

Why it matters: Underfitting (high bias) and overfitting (high variance) are the two ends of the bias-variance tradeoff. Underfitting is less common in practice than overfitting because modern ML methods have high capacity, but it's a real failure mode when the model is too constrained for the problem.

Common mistakes: Assuming that a model that performs poorly on test data is overfitting. If training performance is also poor, the model is underfitting — the fix is more model capacity (deeper tree, more features, more complex model), not regularization.

Related overfitting, decision tree, linear regression, variance