Understanding Training, Testing, Actual & Predicted Values

Apple Example

Think of a Machine Learning Model as a Child

A machine learning model is like a child who does not know anything initially.

Just like a child:

• The model must be taught
• The model must then be tested

Training Phase – Teaching the Child

You start teaching the child using fruits.

You show a fruit and ask:

“What fruit is this?”

The child answers something:

• Maybe “Apple”
• Maybe “Banana”

This answer is called the predicted value
Because it is the child’s guess.

You immediately tell the child:

“The correct answer is APPLE 🍎”

This correct answer is called the actual value.

So during training:

• Child gives a predicted value
• Human provides the actual value
• Child learns by comparing both

This process repeats many times.

What Exactly Is Training Data?

The apples you show during teaching are called training data.

 Training data means:

Data used to teach the model, along with correct answers (actual values)

Error – How Learning Happens

If:

• Predicted value = Apple
• Actual value = Apple
  → No error

If:

• Predicted value = Banana
• Actual value = Apple
  → Error happened

 Error = Difference between predicted and actual value

During training:

• Error is used to improve learning

Testing Phase – Exam Time (New Apple 🍎)

Now the exam starts.

You show a new apple:

• The child has never seen this exact apple
• But it is still an apple

 This new apple is called test data.

You ask:

“What fruit is this?”

The child answers:

“Apple”

 This answer is again a predicted value.(but during testing)

Step 1: What Happens During the Exam?

Even in the exam:

• Humans already know the correct answer
• The fruit is actually an apple

 That correct answer is still called the actual value.

But now:

• We do not tell the child the answer
• We do not help the child learn
• We only observe what the child answers

---

Step 2: Observing One Answer Is Not Enough

Suppose:

• One apple → child says “Apple” ✔

Can we conclude:

“The child has learned perfectly”?

❌ No.

Because:

• Maybe it was luck
• Maybe it remembered only one apple

---

Step 3: Many Exam Questions (Many Test Data)

So in an exam:

• We ask many questions
• Not just one apple

Example:

• Apple 1 → correct
• Apple 2 → wrong
• Apple 3 → correct
• Apple 4 → wrong

 Now the big question:

“Overall, how well did the child perform?”

---

 How Humans Measure Exam Performance

In school:

• We don’t check question by question
• We calculate marks, percentage, grade

 That single score tells:

• How good the performance is
• Whether the child passed or failed

---

 Machines Need the Same Thing

In machine learning:

• The model predicts on many test data points
• For each one:
  • We know the actual value
  • We get a predicted value

 We need:

One or a few numbers to summarize performance

---

 This Summary Is Called a Performance Metric

What Is a Performance Metric? (Slow Introduction)

A performance metric is:

• A numerical measure
• That summarizes how accurate the model is
• Based on errors on test data

 It answers:

• How close are predictions to actual values?
• Is the model reliable?
• Is this model better than another one?

---

 Apple Example → Performance Metric

Actual (Human Truth)	Predicted (Child)
Apple	Apple
Apple	Banana
Apple	Apple
Apple	Apple

Instead of checking one by one, we say:

“Child answered 3 out of 4 correctly.”

That score is like a performance metric.

---

From Apple to Regression

In regression:

• Instead of correct / wrong
• We measure how far the prediction is

Example:

• Actual price = ₹45 lakh
• Predicted = ₹50 lakh
• Error = ₹5 lakh

Performance metrics combine many such errors.

Training vs Testing — Final Clear Explanation

(Learning Phase vs Exam Phase)

---

 Training Phase = Learning Phase

In the training phase, the model is learning, just like a child in class.

What happens?

• The model predicts an answer
• Humans already know the actual value
• The model compares:
  • Predicted value
  • Actual value
• The difference (error) is used to improve learning

 Actual values are used to TEACH the model

Just like:

A teacher corrects the child after each question in class

---

Testing Phase = Exam Phase

In the testing phase, the model is examined, not taught.

What happens?

• The model predicts on new data
• Humans still know the actual value
• BUT:
  • The model is not corrected
  • No learning happens
• We only observe predictions

 Actual values are used only to CHECK performance

Just like:

Teachers already know the answers in exams, but they only calculate marks

---

 Why Performance Metrics Are Calculated on Test Data

Now make this conclusion very explicit:

Performance metrics are calculated on test data, not training data.

Why?

• Training data is for learning
• Test data is for evaluation
• Test data shows real ability

 Same logic as:

• Classwork → learning
• Exam → performance

Final Comparison Table 

Aspect	Training Phase	Testing Phase
Purpose	Learning	Evaluation
Data used	Training data	Test data
Predicted values	Yes	Yes
Actual values used	Yes (for correction)	Yes (for checking only)
Learning happens	Yes	No
Performance metrics	No	Yes

Actual values are used during training to improve the model’s learning, while during testing they are used only to evaluate performance using performance metrics.

In exams,

teachers calculate:

• Marks
• Percentage
• Grade

That single number tells:

• How well the student performed overall

Same Question for Machine Learning Models

After testing the model on many test data points, we ask:

• Is the model good or bad?
• How big are its mistakes?
• Is Model A better than Model B?
• Are mistakes small or dangerous?

 We cannot judge by looking at individual predictions.

This Is Where Regression Metrics Come In

Regression performance metrics are numerical scores that summarize how well a regression model predicts real-world numerical values.

Just like:

• Marks summarize exam performance

Regression metrics summarize:

• Errors made by the model on test data

Why ONLY Test Data? (Clear & Final)

Make this point very clear to avoid confusion:

✔ Training data is used for learning
✔ Test data is used for evaluation
✔ Performance metrics are calculated on test data

Same logic as:

• Classwork → learning
• Exam → marks

House Price – Regression

• Actual house price = ₹45 lakh (human knows)
• Predicted house price = ₹50 lakh (model guesses)
• Error = ₹5 lakh

Now imagine:

• 100 houses
• 100 such predictions
• 100 errors

How do we summarize all these errors?

Using regression metrics like:

• MAE
• RMSE
• R²
• MAPE