Underfitting is the opposite failure to overfitting. An overfit model scores well on its training data and worse elsewhere; an underfit model scores poorly everywhere, because it never learned the pattern in the first place. The textbook Deep Learning (Goodfellow, Bengio and Courville, 2016) defines underfitting as occurring when a model cannot reach a sufficiently low error on the training set.
Low scores do not always point to the model. In CUBIG’s UCI Polish Companies Bankruptcy experiment (October 2026), a random forest with unchanged settings reached an F1 of 0.41 on the baseline training data and 0.61 after additional data preparation (added features and minority-class augmentation), scored on the same 1,942 evaluation records. The write-up does not report training scores, so it does not show that the baseline was underfit, and it does not isolate which step drove the change or test threshold tuning or class weighting. It is a single split.
Common causes:
- a model too simple for the pattern, such as a straight line fit to a curved relationship, the case the scikit-learn underfitting example uses
- inputs that leave out the signal the task depends on, or encode it poorly
- too much regularization, which holds the model back from fitting real structure
- training stopped too early, before the model learned the pattern
How is underfitting different from overfitting?
- Underfitting: low scores on training data and on unseen data. The model has not learned enough.
- Overfitting: high scores on training data and lower scores on unseen data. The model has learned the training data too closely.
- Both are measured by comparing training scores with scores on records the model never saw, such as a holdout set.
Where does underfitting break in AI work?
A team sees poor results and reaches for a bigger model. If the inputs do not carry the information the task needs, a more complex model has little to learn from, and it may start fitting noise instead, trading underfitting for overfitting.
Before changing the model, it is worth checking what the data offers the task: which fields carry the signal, whether rare outcomes have enough examples, and whether the evaluation set reflects the cases that matter. Comparing a fixed model on the original and prepared data, under the same evaluation conditions, shows whether preparing the data changes the result for that model, though not which change caused it.