Underfitting

Underfitting describes failure to capture relevant structure adequately for a prediction task, often associated with insufficient flexibility or excessive regularization. Similar performance problems can also arise from features, data, objectives, or optimization.

Underfitting occurs when a model cannot adequately capture task-relevant structure in the data under the chosen representation and fitting procedure. Insufficient flexibility or excessive regularization can contribute, but poor performance can also arise from missing features, unsuitable objectives, data problems, or incomplete optimization. A simple model is not underfitted merely because it is simple.

Compare training and validation behavior with appropriate baselines and residual patterns. A persistent systematic error suggests structure the model misses, while adding flexibility can also fit noise. The aim is a useful generalization under the intended conditions, not the most elaborate representation. Outside statistics, an oversimplified explanation is an analogy that should be tested against concrete failures.

When to use it

When a model shows persistent systematic errors or performs poorly on both fitting and validation data and the reason needs investigation.

How it can help

Inspect training and validation behavior, compare baselines, and test targeted changes against independent data rather than increasing complexity indiscriminately.