Underfitting happens when a machine learning model is too simple, or not trained enough, to capture the real patterns in the data. It performs poorly on the training data and on new data. Where an overfit model memorises too much, an underfit model learns too little.
Key takeaways
- The main sign is low performance on both training and validation data.
- Common causes are overly simple models, weak features, too much regularisation and too little training.
- Fixes include more complex models, better features and longer training.
- Underfitting is linked to high model bias.
Underfitting explained in simple terms
Imagine predicting ice cream sales using only the rule “sell 100 cones a day”. It ignores weather, weekends and holidays. It will be wrong on hot days, cold days and most days in between, both in past records and in the future. The rule is too simple for the real pattern. That is underfitting.
Visually, if the data follows a curve and you fit a straight line through it, the line misses the shape everywhere. No amount of extra data fixes that, because the model cannot represent a curve.
How to spot underfitting
| Signal | What it suggests |
|---|---|
| Low training score and similar low validation score | Classic underfitting |
| Barely beats a simple baseline | Model is not capturing useful patterns |
| Training loss stays high and flat | Model lacks capacity or training stopped too early |
| Errors follow a clear pattern | Structure in the data is being missed |
What causes underfitting?
- Model too simple, such as linear regression on a curved relationship.
- Weak or missing features that do not carry enough information.
- Too much regularisation, which forces the model to be overly simple.
- Too little training, such as stopping a neural network after very few epochs.
- Learning rate problems, where gradient descent takes steps too small to make progress.
- Poor data quality, where noise hides the signal.
How to fix underfitting
| Fix | How it helps |
|---|---|
| Use a more flexible model | Move from linear models to trees, boosting or neural networks |
| Better feature engineering | Add informative features, interactions and transformations |
| Reduce regularisation | Allow the model more freedom to fit patterns |
| Train longer | Give the optimiser time to reduce the loss |
| Tune the learning rate | Help optimisation make meaningful progress |
| Add model capacity | More layers, units or deeper trees |
Note that adding more data usually does not fix underfitting. More data helps with overfitting; underfitting needs a more capable model or better features.
Underfitting vs overfitting
| Aspect | Underfitting | Overfitting |
|---|---|---|
| Training performance | Poor | Very good |
| Validation performance | Poor | Much worse than training |
| Linked to | High bias | High variance |
| Main fix | More capacity, better features | More data, regularisation, simpler model |
Finding the right balance
The goal is a model that fits real patterns without fitting noise. A practical approach: start simple as a baseline, increase complexity step by step, and watch the validation score. Stop when validation performance stops improving. That point marks the balance between underfitting and overfitting.
Why it matters
Underfit models waste the value in your data. They are often safe and stable, but they leave accuracy, revenue and insight on the table. In business, a model that barely beats “same as last month” may not justify its cost.