What is underfitting?
Underfitting is the failure mode in which a model is too simple, too constrained, or too poorly fed to learn the pattern it is meant to learn. The sign is poor performance on the training data itself. An overfit model at least gets the training data right. An underfit model gets both the training data and the new data wrong by a similar margin.
Common causes include a linear model applied to a relationship that curves, a tree that was cut off too early, regularization set too strong, too few training passes, or input features that do not contain the needed information. A model predicting store sales from the day of the week alone will underfit if weather and promotions drive most of the variation.
Why underfitting matters
Underfitting is cheaper to detect than overfitting because it shows up before deployment. The risk is misdiagnosis. Teams sometimes respond to a weak model by collecting more of the same data, which does not help when the model lacks the capacity or the features to use it. A retailer whose demand forecast misses every holiday spike needs a holiday feature or a model that can represent spikes, and more years of history alone will not fix it.
The cure is usually more capacity or better inputs: a more flexible model architecture, feature engineering that exposes the relevant signal, weaker regularization, or longer training. The bias-variance tradeoff describes the balance between this problem and its opposite.