In many machine‑learning models—especially logistic regression—we make predictions that look like a score between 0 and 1. This score can be interpreted as the model’s confidence that something belongs to the “positive” class. But the score itself is not the final decision.
To transform this score into a clear answer (Positive or Negative), we use something called a decision threshold.
The threshold is a number, usually 0.5, that acts like a gate:
If model score ≥ threshold → predict Positive
If model score < threshold → predict Negative
And this single number has a huge effect on how the model behaves.
Below we will explore both cases—low threshold and high threshold—and explain not only what happens, but also why each choice is useful in real life.
Imagine a magical “YES detector.”
You ask it: “Is this object blue?”
The machine doesn’t answer yes/no. Instead, it says a number between 0 and 1:
- 0 → “Definitely not blue.”
- 1 → “Definitely blue.”
- Numbers in the middle → “Hmm… not sure!”
You must choose a rule—your threshold—to decide at what point the machine’s score counts as “blue.”
If your threshold is low, you accept almost any weak confirmation.
If your threshold is high, you only accept very strong confirmation.
This is the whole game.
Low Threshold: High Recall, Low Precision
A low threshold means the model requires very little confidence to call something “Positive.”
High Recall
Recall answers:
“Out of all the real positives, how many did the model catch?”
With a low threshold, the model labels many things as Positive—even ones with weak evidence (FP).
So it rarely misses real positives.
This means very few false negatives (FN) : recall becomes high.
Low Precision
Precision answers:
“Of all the predicted positives, how many were truly positive?”
Because the model is so eager to say “Positive,” it accidentally includes many negatives.
This increases false positives (FP) : precision becomes low.
Friendly Analogy
Imagine using a huge fishing net to catch butterflies.
You catch almost all butterflies (high recall)…
…but also catch some leaves and rocks (low precision).
Practical Use Cases for Low Threshold
Low thresholds are used when missing a positive case would be dangerous, for example:
- Medical screenings (early detection of disease)
- Safety alerts (fire alarms, gas leak alerts)
- Fraud detection
- Early pregnancy detection
Here, it is better to say “maybe positive” too often than to miss a true positive.
High Threshold : High Precision, Low Recall
A high threshold means the model requires very strong confidence to call something “Positive.”
High Precision
Because the threshold is high, only the very clear positive cases are accepted.
Wrong positives (FP) become rare : precision becomes high.
Low Recall
But many real positives do not reach this high confidence bar.
They fall below the threshold and are wrongly classified as Negative.
These become false negatives, making recall lower.
Friendly Analogy
Now imagine catching butterflies using a tiny little net.
The ones you catch are definitely butterflies (high precision)…
…but you miss many (low recall).
Practical Use Cases for High Threshold
High thresholds are used when false alarms are expensive or harmful, such as:
- Spam detection (you don’t want real emails marked as spam)
- Loan approval systems (better to reject doubtful cases)
- Security checkpoints (avoiding unnecessary interventions)
- Final medical diagnosis (confirmation after preliminary tests)
Here, being very sure before saying “Positive” is the priority.
Changing the threshold is like adjusting a dial between:
- Catching as many positives as possible (high recall)
- Being very accurate when predicting positives (high precision)
You cannot have both high recall and high precision at the same time using a single threshold.
The threshold simply tells the model which type of mistake you prefer to avoid:
- Low threshold : avoid missing positives (avoid FN)
- High threshold : avoid false alarms (avoid FP)
Additional Resources
How to use classification threshold to balance precision and recal
Clasificación: Exactitud, recuperación, precisión y métricas relacionadas
Calculating and Setting Thresholds to Optimise Logistic Regression Performance
© Image. https://www.evidentlyai.com/
@Yolanda Muriel 
yolandamuriel.com 