Underfitting
Underfitting is a model too simple, too constrained or too briefly trained to capture the structure in its data, so that it performs poorly on the training set and on unseen data alike — the opposite failure to overfitting, where training performance is high and generalisation poor. Its signs are a training error that stays high and a small training–validation gap; its causes are too few parameters, too much regularisation, too few steps, or features that miss the signal.
- Also known as
- Underfit, High bias, Model too simple
- Used at
- model development, where the training curve is read for both failures; and model review, where a model that underfits misses risk the data could have revealed
- Binding
- none; but a credit or fraud model that underfits leaves detectable risk undetected, which a reviewer will call a weakness in the same way as an overfitted one
- Typical signs
- high training loss, training and validation metrics close together and both poor, a learning curve that has flattened at an unsatisfactory level, residuals with visible structure (the model systematically misses a segment)
- Typical causes
- a linear model for a non-linear relation, trees too shallow, a network too small, regularisation too strong, early stopping too early, a learning rate too low, insufficient epochs, or features that do not encode what matters
- Typical remedies
- add capacity (depth, units, trees), reduce regularisation, train longer, engineer or add features, or change model class; then re-check for overfitting, which is what appears when capacity is added without care
- Parties
- data scientists reading learning curves; reviewers asking whether the model is as good as the data allows
How underfitting works in practice
A fraud model built as a logistic regression on 30 hand-picked features scores an AUC of 0.78 on training and 0.77 on validation. The gap is small — no overfitting — but a gradient-boosted model on the same features scores 0.88 on both; the regression could not represent the interactions (this merchant category at this hour on this device type) that the trees find. The regression underfit: not because it memorised noise but because it could not express the pattern. Adding capacity closed the gap; adding capacity carelessly (depth 15, no regularisation) then produced overfitting, and the tuned model at depth 6 landed at 0.90 validation. Reading the two failures together is the daily work: the learning curve says which side of the optimum the model sits on.
Underfitting hides more easily than overfitting because it looks stable. A model whose training and validation scores agree is reassuring; if both are mediocre, the question is whether the data could support better. The diagnostic is to try a more flexible model, or a larger one, or more features, and see whether validation improves. Features are the usual answer in tabular finance: a model without velocity features cannot detect velocity fraud however deep its trees, and a credit model without affordability features underfits the affordability risk. In deep learning the common causes are a learning rate too low, training stopped too soon, or regularisation copied from a larger dataset's recipe.
For a fintech, underfitting is the quieter of the two errors and the costlier when it hides: the fraud not caught and the default not predicted were in the data, and the model was too simple to see them.
Example
A lender's affordability model, a scorecard with 12 binned variables built to be explainable, reaches a Gini of 0.41 on both development and validation. The model-risk team notes the agreement and passes it. A challenger built by a new analyst — gradient boosting on 60 features including bureau-derived affordability ratios — reaches 0.58 on validation and 0.57 out of time. The scorecard was underfitting: too few, too coarse variables for the relation in the data. The lender adopts the challenger with SHAP-based reason codes to preserve explainability, and the incremental prediction on a €300 million book is worth about €2.1 million a year in avoided losses at constant approval rates — risk that had been in the data all along.
Common mistakes
- Reading a small training–validation gap as proof the model is good; it may be equally bad on both. - Regularisation copied from another problem and never swept. - Stopping training when the loss is still falling. - A linear model for a relation the data shows is not. - Fixing underfitting by adding capacity without re-checking for overfitting.