Short Definition

Calibration vs Raw Accuracy contrasts two evaluation dimensions of machine learning models: Raw Accuracy measures how often predictions are correct, while Calibration measures how well predicted probabilities reflect true outcome frequencies.

A model can be accurate but poorly calibrated.

Definition

In classification tasks, models output:

• A predicted class
• A confidence score (probability)

Two distinct properties matter:

Raw Accuracy

[
\text{Accuracy} = \frac{\text{Correct Predictions}}{\text{Total Predictions}}
]

Accuracy evaluates correctness of final decisions.

It does not assess whether confidence scores are trustworthy.

Calibration

A model is calibrated if:

[
\mathbb{P}(Y = \hat{Y} \mid \hat{p} = p) = p
]

Meaning:

If the model predicts 70% confidence,
it should be correct about 70% of the time.

Calibration evaluates probability reliability.

Core Difference

Aspect	Raw Accuracy	Calibration
Measures	Correctness	Confidence reliability
Concern	Classification outcome	Probability trustworthiness
Sensitive to	Decision threshold	Confidence distribution
Risk if poor	Wrong answers	Misleading confidence

Accuracy answers: “Was it right?”
Calibration answers: “Can I trust its confidence?”

Minimal Conceptual Illustration

“`text
Model A:
Accuracy = 90%
Confidence always = 99%
→ Overconfident.

Model B:
Accuracy = 90%
Confidence ≈ true correctness rate
→ Well calibrated.

Both equally accurate.
Only one trustworthy.

Why Accuracy Is Not Enough

High accuracy does not guarantee:

• Reliable uncertainty estimates
• Safe deployment decisions
• Appropriate thresholding

In safety-critical systems, confidence calibration matters as much as correctness.

Common Miscalibration Patterns

1. Overconfidence
   • Predicted probabilities too high.
   • Common in deep neural networks.
2. Underconfidence
   • Predictions overly cautious.
3. Confidence collapse
   • Overconfidence under distribution shift.

Modern neural networks tend to be overconfident.

Measuring Calibration

Common metrics:

• Expected Calibration Error (ECE)
• Maximum Calibration Error (MCE)
• Brier Score
• Reliability Diagrams

Calibration is typically evaluated via probability binning.

Trade-Off Dynamics

Improving accuracy does not necessarily improve calibration.

In fact:

• Larger models often increase confidence.
• Scaling can worsen miscalibration.
• Regularization affects calibration behavior.

Calibration requires explicit evaluation.

Thresholding Implications

Many systems use confidence thresholds:$$\text{Accept prediction if } p > \tau$$Accept prediction if p>τ

If poorly calibrated:

• High confidence may be misleading.
• Risk-sensitive decisions become unsafe.

Calibration affects operating point selection.

Distribution Shift Effects

Under distribution shift:

• Accuracy often drops.
• Confidence may remain high.
• Miscalibration increases.

Robust systems must monitor calibration drift.

Alignment Perspective

Poor calibration can lead to:

• Overconfident hallucinations.
• Misleading outputs.
• Unsafe automation decisions.

Alignment requires not only correctness,
but truthful confidence estimation.

Calibration is central to trustworthy AI.

---

Governance Perspective

In regulated domains:

• Medical diagnosis
• Autonomous systems
• Financial decision systems

Confidence must reflect risk.

Policies often require:

• Probability reliability auditing
• Ongoing calibration monitoring
• Threshold governance

Calibration is part of risk management.

Improving Calibration

Common techniques:

• Temperature Scaling
• Label Smoothing
• Mixup
• Bayesian methods
• Ensemble averaging

Calibration is typically applied post-training.

Summary

Raw Accuracy:

• Measures correctness.
• Ignores probability reliability.

Calibration:

• Measures trustworthiness of confidence.
• Critical for risk-sensitive decisions.

High accuracy without calibration can be dangerous.

Reliable AI systems require both.

Related Concepts

• Calibration
• Reliability Diagrams
• Expected Calibration Error (ECE)
• Decision Thresholding
• Operating Point Selection
• Confidence Collapse
• Distribution Shift
• Uncertainty Estimation