Underfitting

Underfitting is the condition where a model fails to capture the patterns present in the training data itself — training error is high, test error is high, and more training does not help because the hypothesis space the model can represent is simply too small for the problem. A linear regression on clearly nonlinear data, a shallow decision tree on a problem that needs depth, a tiny neural network on a complex task — all classic underfitting. The telltale sign is that training accuracy plateaus far below what the problem ought to allow.

Underfitting sits opposite overfitting on the bias–variance tradeoff. The cures are the mirror image: increase model capacity (more parameters, more layers, higher tree depth, richer features), remove over-aggressive regularisation (lower weight decay, drop the dropout rate, loosen the L2 penalty), train longer, or add more informative features. The diagnostic that separates it from other failures is that both training and validation error stay high together; if training error is low but validation error is high, that is overfitting.

In modern production ML on APAC mid-market stacks, underfitting is less common than overfitting because the default toolchains (XGBoost, LightGBM, large pretrained transformers) are extremely expressive. When it does appear, it is usually a signal that the feature set is wrong — the model cannot learn something it cannot see. Adding raw model capacity rarely fixes that; adding the missing signal does.

One non-obvious form: **distributional underfitting**, where average-case metrics look fine but the tails are bad because the dominant class drowns out rare-but-important patterns. Subgroup-sliced evaluation surfaces this before it ships.