Underfitting
Last updated June 11, 2026
What is Underfitting in simple terms?
In simple terms, underfitting is when a model is too simple to get it. It misses the real pattern and does badly even on the examples it studied — like skimming one page before an exam and grasping almost nothing.
What is Underfitting?
Underfitting is when a machine learning model is too simple to capture the real patterns in its training data, so it performs poorly even on the examples it was trained on, let alone on new ones.
Underfitting is the failure that happens when a model is too simple for the job. The real pattern in the data is richer than the model can represent, so it never really learns the relationship — it does poorly not just on new data, but even on the very examples it was trained on. That last part is the tell-tale sign: an underfitting model is bad across the board, because it hasn't captured enough of the underlying structure to do well anywhere. It's like trying to summarize a whole book from a single page — there simply isn't enough being taken in to get it right.
Underfitting usually comes from a model that lacks the capacity or information to do better. Maybe it's too basic for a genuinely complex problem, maybe it was given too few useful features to work with, maybe it wasn't trained long enough, or maybe its complexity was reined in too aggressively in an effort to keep it simple. Whatever the cause, the result is a model that produces crude, low-accuracy outputs — often defaulting to something close to a flat average — because it's blind to the real patterns that would let it do better. More data won't fix it on its own; the model itself needs more power or better inputs.
Underfitting is the opposite failure to overfitting, and together they frame the central balancing act of machine learning. Overfitting is a model that memorized its training data and can't generalize; underfitting is a model too simple to learn the data in the first place. The goal sits between them: a model complex enough to capture the genuine pattern but not so complex that it memorizes noise — the sweet spot of good generalization. Spotting underfitting is usually straightforward, since poor performance even on training data is a clear signal, and the fix is to give the model more capacity, better features, or more training.
Real-world example of Underfitting
An ice-cream van owner wants to predict daily sales so she can stock the right amount, and a friend builds her a model — but it uses only one piece of information: whether the day is a weekend. The model is far too simple for the problem. Sales really depend on temperature, the season, school holidays, and the weather, none of which it can see, so it can't help but be crude: it basically learns 'sell a bit more at weekends' and otherwise guesses near the average. It's wrong on hot Tuesdays and cool Saturdays alike — and tellingly, it's wrong even on the past days it was trained on, because one weekend-or-not flag simply can't capture what drives ice-cream sales. That's underfitting: the fix isn't more of the same data but a richer model fed the factors that actually matter.
Related terms
Frequently asked questions about Underfitting
What is the difference between underfitting and overfitting?
They're opposite failures. Underfitting is when a model is too simple to capture the real pattern, so it does poorly even on its training data. Overfitting is when a model is too complex and memorizes its training data, including random noise, so it does great on what it studied but poorly on anything new. Underfitting means the model hasn't learned enough; overfitting means it has learned too much of the wrong thing. The goal is the balance between them — a model that learns the genuine pattern without memorizing the specifics.
How do you fix underfitting?
Because underfitting comes from a model that's too simple or starved of useful information, the fixes add capability rather than restrict it. You can use a more powerful model that can represent more complex patterns, give it better or additional input features so it has more to learn from, train it for longer, or ease off any constraints that were forcing it to stay overly simple. The clearest sign you've succeeded is that performance improves on the training data itself — since underfitting shows up as poor results even there, not just on new data.
Why does underfitting matter?
It matters because an underfitting model is essentially useless — it performs poorly everywhere, so it can't be trusted to make good predictions. Recognizing it is important because the remedy is the opposite of the remedy for overfitting: underfitting calls for a more capable model or richer inputs, while overfitting calls for simplifying or constraining the model. Diagnosing which problem you have prevents you from making things worse, and understanding underfitting is part of finding the balance point where a model generalizes well to new data.