Not sure which lesson to do first? Find your level and get a plan →

Math · Statistics & Probability · Grade 9-12

Underfitting (Intuition)

Underfitting happens when a model is too simple to capture the true pattern, so it performs poorly on both training data and new data.


Quick Answer

Underfitting happens when a model is too simple to capture the true pattern, so it performs poorly on both training data and new data. Use the idea when a model does badly even on the data it was built from: a straight line through data that clearly curves, or house prices predicted from the number of bedrooms alone. Because the training and new-data scores are similar, the problem is missing structure, not memorization, and the fix is a richer model: a curve, more input variables, or longer training.

Why This Matters

Underfitting wastes the information in your data. A shop that forecasts ice cream sales with one flat yearly average misses the summer peak and runs out every July. Recognizing underfitting tells you to make the model more flexible rather than collect more of the same data, which cannot help a model that is unable to bend.

Intuitive Explanation

Picture data shaped like a smile: high on the left, low in the middle, high on the right. Fit a straight line and the best it can do is run fairly flat through the middle. It predicts too low at both ends and too high in the middle, and no choice of slope or intercept fixes that, because a line cannot bend. The model is too simple for the pattern, and that is underfitting.

The scores give it away. An underfit model does poorly on its training data, say 55%, and about equally poorly on new data, say 54%. There is no gap, because the model memorized nothing; it never captured the pattern in the first place. Its residuals have a shape too: for the smile they are positive at both ends and negative in the middle. Switch to a quadratic, which can bend, and both scores should rise together.

Low scores on new data alone do not mean underfitting. A model that scores 99% on training but 61% on new data learned too much, not too little: that is overfitting, and its cure is the opposite, a simpler model. Underfitting is also different from noisy data. If the data itself is very random, even the right model leaves large residuals, but those residuals scatter without any pattern.

Think of it like this. An underfit model is like describing a mountain range as 'about 2,000 metres high'. The number is not crazy, but it misses every peak and valley, and repeating it more confidently does not help. You need more detail, such as a contour map. The comparison has a limit built in: too much detail, mapping every pebble, becomes the opposite problem, overfitting.

Core idea. Underfitting is when a model is so basic it misses the real structure and does poorly everywhere.


When to Use

Suspect underfitting when a model does badly on its own training data as well as on new data, with no big gap between the two, or when its residuals show a clear curve or trend. Common causes are a straight line for curved data and too few input variables. If training scores are high and new-data scores much lower, that is overfitting instead; if residuals are large but patternless, the data may simply be noisy.

Tip. Ask: Does the model do badly even on the data it was trained on?

How to Recognize It

Look at how the model does on the very data it was built from, and at whether its misses form a shape or scatter at random.

  1. Are the training error and the new-data error both high and close to each other?

    Yes points to underfitting: the model is too simple to capture the pattern anywhere. If the training error is low and only the new-data error is high, the model is overfit instead.

  2. Do the residuals form a curve or a trend, such as positive at both ends and negative in the middle?

    Yes means real shape is being missed, so a curved model or an extra input may help. Large residuals with no shape at all suggest noisy measurements rather than a model that is too simple.

  3. Is the model using fewer inputs than clearly affect the outcome?

    Yes is a common cause: predicting a car's fuel use from its weight alone ignores engine size and speed, so even the training data fit poorly. Adding the missing variables is the fix, not collecting more rows of the same two columns.

  4. Do the data points scatter widely even when the inputs are identical?

    Yes means some error cannot be removed by any model: two students with the same study hours can get different scores. That spread is measurement error or natural variation, so judge the model against it rather than against zero.

Underfitting (Intuition) compared with Overfitting, Model fit (good), and Measurement error

These rows are all reasons predictions miss. The separating question is where the misses happen: on training and new data alike (underfitting), only on new data (overfitting), on neither (good fit), or because the measurements themselves are unreliable (measurement error).

Underfitting (Intuition)

Meaning
Use this when a model performs poorly even on its own training data because it is too simple.
Key test
Does the model do badly even on the data it was trained on?
Example
A straight line is fit to data shaped like a parabola. It scores 55% on training and 54% on new data. What's wrong?

Overfitting

Meaning
The opposite failure: model too complex, memorizes noise, fails only on new data.
Key test
Use when the model is great on training but poor on new data.
Example
A wiggly curve through every training point

Model fit (good)

Meaning
Captures the real pattern and does well on both training and new data.
Key test
Use for the healthy middle between the two failures.
Example
Curve follows the trend without chasing noise

Measurement error

Meaning
Noise in how data was recorded, not a flaw in the model.
Key test
Use when the data itself is off, not when the model is too simple.
Example
A miscalibrated scale reads 2 kg high

Worked Examples

Example 1 — Spot the underfit model

Easy

A straight line is fit to data shaped like a parabola. It scores 55% on training and 54% on new data. What's wrong?
Show the steps and the answer
  1. Find the gap: 55%−54%=1 percentage point.

    With almost no gap, the model has not memorized its training data, so overfitting is ruled out.

  2. Look at the level: 55% on the very data it was fit to.

    A model that cannot fit its own training data is missing the pattern.

  3. Compare shapes: a line cannot bend, but parabola-shaped data does.

    Whatever slope is chosen, the residuals will have the same sign at both ends and the opposite sign in the middle.

  4. Fix it with a model that can bend, such as a quadratic.

    Matching the model's shape to the data's shape should raise both scores together.

Answer: Underfitting

Takeaway: Poor scores on both sets with little gap mean underfitting. Give the model the shape the data needs.

Example 2 — Memorizing, not missing

Standard

A model scores 99% on training but 61% on new data. Is it underfit?
Show the steps and the answer
  1. Check the training score: 99%.

    An underfit model does poorly on its training data, and this one nearly memorized it.

  2. Find the gap: 99%−61%=38 percentage points.

    A big drop on new data is the signature of overfitting.

  3. Choose the fix: simplify the model or train it on more data.

    Making the model richer, the cure for underfitting, would widen the gap further.

Answer: Overfit, not underfit

Takeaway: Poor new-data scores alone do not mean underfitting. A high training score with a big drop is overfitting.

Example 3 — Find the mistake: simplifying an underfit model

Application

A model predicts a car's stopping distance from its speed with a straight line. It is off by about 8 m on average on its training data and about 8 m on new data. Stopping distance actually grows roughly with the square of speed. Omar writes: "The model is inaccurate, so it must be overfitting. I'll make it simpler by predicting the same distance for every speed." Find the mistake and correct it.
Show the steps and the answer
  1. Compare the two errors: about 8 m on training data and 8 m on new data.

    Overfitting needs a much smaller training error than new-data error, and here there is no gap.

  2. Compare shapes: braking distance grows with the square of speed, so the data curves upward, but a line cannot curve.

    A line that is too simple for the pattern misses the same way on every data set, which is underfitting.

  3. Test Omar's fix: one constant distance for every speed is even simpler than a line.

    Simplifying an underfit model throws away even the slope it had, so the errors would grow.

  4. Correct the fix: use a model with a squared term, such as d=av+bv2, and check that both errors fall.

    The v term can cover the distance travelled while reacting and the v2 term the braking, so the model can follow the curve.

Answer: The model is underfitting, not overfitting: its errors are equally large on both sets. It needs a richer model with a v2 term, not a simpler one.

Takeaway: Before simplifying a model, compare its training and new-data errors. Equal, large errors call for more flexibility.

Where You'll See It

  • Stopping distances in the Highway Code

    The UK Highway Code lists stopping distances of 12 m at 20 mph, 36 m at 40 mph and 96 m at 70 mph. The best straight line through its six values misses by about 5 m at both ends and 4 m in the middle, in a U pattern, and even predicts a negative distance at 10 mph. A model with a speed-squared term follows the values closely.

  • Forecasting ice cream sales

    A shop that forecasts the same 200 cones every day, its yearly average, is underfitting: it is too low all summer and too high all winter, on past and future data alike. Adding the month or the temperature as an input lets the forecast follow the seasons.

  • Grading essays by length alone

    A rule that scores essays only by word count is too simple to capture quality. It does about equally badly on the essays it was built from and on new ones, because length is only a small part of what makes writing good.

Common Mistakes

  1. The slip: Assuming poor accuracy always means overfitting

    Instead: if training accuracy is also poor, it is underfitting.

  2. The slip: Keeping a model simple to avoid overfitting and ignoring that it never fit at all

    Instead: too simple is its own failure.

  3. The slip: Blaming the data when the model can't even match training points

    Instead: a too-simple model misses real structure that is genuinely there.


Mini Practice

Try these on your own. Tap Reveal when you want to check.

  1. Model 1 scores 62% on training data and 60% on new data. A more flexible Model 2 scores 88% and 86%. Which model was underfit?

    Hint: An underfit model does poorly even on the data it learned from.

  2. For x=1,2,3,4,5 the data are y=1,4,9,16,25. A flat line y=11 is used as the model. Find the residuals.

    Hint: Residual means actual y minus 11.

  3. A straight line is fit to y=x2 for x=−3,−2,−1,0,1,2,3. The best line is y=4. Find the residuals at x=0 and x=3.

    Hint: Compute x2 at each point, then subtract 4.

  4. Models of degree 1,2,6 have training errors 40,12,2 and test errors 41,13,30. Which model underfits, which overfits, and which is best?

    Hint: Underfit is high on both; overfit is low on training but high on test.

  5. A model's errors on training and new data are all within ±0.5 kg, about the same as the scale's own measurement error. Is the model underfit, or is this measurement error?

    Hint: Can any model be more accurate than the measurements themselves?

  6. A model gets 55% on training data and 53% on new data. Ria says: "It is overfit, so we should make it simpler." Correct her.

    Hint: Is there a big gap between training and new data?

Want the full set?

50 practice questions for this concept — free to try, every one with a complete worked solution showing the why, not just the answer.

Frequently Asked Questions

Can more data fix an underfit model?

Usually not. A straight line through curved data stays a straight line whether you give it 50 points or 50,000; extra points only pin down the wrong shape more precisely. More data is the cure for overfitting. For underfitting you change the model itself: add a squared term, add the input variables that were left out, or, for a learning system, train it for longer.

Why is underfitting sometimes called high bias?

Bias here means the model's built-in assumption pulls its predictions away from the truth in the same way every time. A straight-line model assumes the relationship is straight, so on curved data it makes the same kind of miss on every sample you could draw: too high in one region, too low in another. Overfitting is the opposite, high variance, where predictions jump around depending on which sample the model saw.

Is the average of all the y values a model?

Yes, and it is the simplest one: predict yˉ\bar{y} for everyone, whatever xx is. It is the usual baseline for spotting underfitting. R2R^2 actually measures how much better a model does than this flat line, so an R2R^2 near 0 means your model has learned about as little as predicting the mean every time would.

Learning Path

Underfitting (Intuition)

You are here

Next →

Nothing after this yet
You need Model Fit (Intuition) first: residuals, and how to tell a good fit from a poor one. Underfitting names the failure where a model is too simple to capture the pattern, and shows how to fix it. Together with overfitting, it frames every choice of model complexity, from picking linear, quadratic or exponential models to tuning a predictive model on held-out data.

See Also