Lesson 4 of 7 · 11 min
Overfitting, underfitting and model fit
A useful ML model learns the true signal: overfit models memorise noise and fail on new data, underfit models miss real patterns, and comparing training with test performance tells you which you have.
In short
- Overfitting: the model learns the training data too precisely and treats noise as true parameters; it is often too complex and predicts poorly on a different dataset.
- Underfitting: the model treats true parameters as noise and misses real relationships; it is often too simplistic.
- Signature of overfitting: excellent fit on training data, much worse on test data.
- Signature of underfitting: weak fit on training data and on test data.
- Overfitted models find false or unsubstantiated patterns that cause prediction errors and wrong forecasts.
Unlock this lesson free for 7 days
Create a free account to get 7 days of full access — every lesson, video, flashcard, mock and the question bank. No card needed.