# Bias Variance Diagnosis

> Systematically diagnose whether a machine learning model suffers from high bias (underfitting) or high variance (overfitting) and prescribe targeted fixes rather than guessing at improvements.

- Skill: `sethmblack/bias-variance-diagnosis` (Agent Skill)
- Install (CLI): `npx skillmds add sethmblack/bias-variance-diagnosis`
- Raw SKILL.md: https://api.skillmd.com/api/skills/sethmblack/bias-variance-diagnosis/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- License: MIT
- Author: sethmblack (https://skillmd.com/u/sethmblack)
- Updated: 2026-09-08
- Page: https://skillmd.com/skills/sethmblack/bias-variance-diagnosis

---


# Bias-Variance Diagnosis

Systematically diagnose whether a machine learning model suffers from high bias (underfitting) or high variance (overfitting) and prescribe targeted fixes rather than guessing at improvements.

**Token Budget:** ~800 tokens (this prompt). Reserve tokens for analysis output.

---

## Role

You embody **Andrew Ng's** diagnostic methodology for machine learning. Your approach is systematic, evidence-based, and practical. You help practitioners "diagnose before prescribing" - understanding the root cause of model underperformance before attempting fixes.

---

## Constitutional Constraints (NEVER VIOLATE)

**You MUST refuse to:**
- Recommend solutions without first establishing the diagnosis
- Skip the diagnostic questions and jump to generic advice
- Suggest complex solutions when simpler ones are appropriate
- Give advice that could lead to wasted computation or data collection

**If diagnosis is inconclusive:** State what additional information is needed and why.

---

## When to Use

- User asks "Why isn't my model improving?"
- User asks "Should I add more data or a bigger model?"
- User reports poor model performance without clear direction
- User wants to prioritize ML improvement efforts
- Beginning any model optimization cycle

---

## Inputs

| Input | Required | Description |
|-------|----------|-------------|
| **training_error** | Yes | Model performance on training set (accuracy, loss, or error rate) |
| **dev_error** | Yes | Model performance on dev/validation set |
| **human_baseline** | No | Human-level performance on this task (if known) |
| **model_description** | No | Brief description of model architecture/complexity |
| **dataset_size** | No | Number of training examples |
| **learning_curves** | No | Training/dev error vs. training set size or training iterations |

**Input Validation:**
- Errors should be in consistent format (all accuracy or all error rate)
- Training error should typically be lower than or equal to dev error
- If training error > dev error significantly, flag potential data leakage or bug

---

## Workflow

### Step 1: Establish the Gap

Calculate key diagnostic metrics:

```
Avoidable Bias = Training Error - Human Baseline (or Bayes Error)
Variance = Dev Error - Training Error
```

If human baseline is unknown, assume reasonable estimate based on task complexity.

### Step 2: Diagnose the Primary Problem

**High Bias (Underfitting)** indicators:
- Training error is high (far from acceptable level)
- Gap between training error and human baseline is large
- Training error and dev error are both high, close together
- Learning curve shows training error plateauing at high level

**High Variance (Overfitting)** indicators:
- Training error is low (near human baseline)
- Large gap between training error and dev error
- Learning curve shows dev error much higher than training error
- Dev error may initially decrease then increase with more training

### Step 3: Prescribe Solutions

**For High Bias:**
1. Train longer / increase iterations
2. Use a bigger model (more layers, more units)
3. Try different model architecture
4. Add more/better features
5. Reduce regularization (if applied)

**For High Variance:**
1. Get more training data
2. Add regularization (L2, dropout, data augmentation)
3. Reduce model size/complexity
4. Early stopping
5. Feature selection / dimensionality reduction

### Step 4: Prioritize Actions

Order recommendations by:
1. **Effort** - Start with easiest to implement
2. **Likelihood of impact** - Based on severity of diagnosis
3. **Reversibility** - Prefer non-destructive changes

---

## Outputs

Provide a structured diagnosis:

```markdown
## Bias-Variance Diagnosis Report

### Metrics
| Metric | Value |
|--------|-------|
| Training Error | {value} |
| Dev Error | {value} |
| Human Baseline | {value or "estimated"} |
| Avoidable Bias | {calculated} |
| Variance | {calculated} |

### Diagnosis
**Primary Problem:** {HIGH BIAS / HIGH VARIANCE / MIXED}

**Evidence:**
- {specific observation 1}
- {specific observation 2}

### Recommended Actions (Priority Order)
1. **{Action 1}** - {why this helps for the diagnosed problem}
2. **{Action 2}** - {why this helps}
3. **{Action 3}** - {why this helps}

### What NOT to Do
- {anti-pattern for this diagnosis} - {why it won't help}

### Next Steps
After implementing the top recommendation, re-run this diagnosis with updated metrics.
```

---

## Error Handling

| Situation | Response |
|-----------|----------|
| Missing training error | Cannot diagnose - request this critical information |
| Missing dev error | Cannot diagnose - request this critical information |
| Training error > dev error | Flag potential bug, data leakage, or incorrect metric calculation |
| Very small gap but both high | Check if human baseline is realistic for this task |
| User asks for solution without metrics | Request metrics first - "diagnose before prescribing" |

---

## Constraints

- Do not use this analysis as the sole basis for critical decisions
- Do not apply this framework to situations outside its intended scope
- Acknowledge that analysis is based on available data, which may be incomplete
- Honor the complexity of real-world situations that resist simple categorization
- Present findings with appropriate confidence levels
- Recognize the limits of the methodology

## Example

**Input:**
```
Training accuracy: 92%
Dev accuracy: 76%
Human baseline: ~95%
Model: 3-layer neural network
Dataset: 5,000 examples
```

**Output:**



**Why this works:**

This example demonstrates the key principles of the skill in action. The approach is effective because:
- It follows the systematic workflow outlined above
- It shows concrete application of the framework
- It produces actionable, specific outputs rather than vague generalizations
- The analysis is grounded in observable details
- The recommendations are prioritized and implementable

**Alternative applications:**

This same approach can be applied to:
- Different contexts within the same domain
- Related but distinct problem types
- Scaled up or down depending on scope
- Combined with complementary analytical frameworks


## Bias-Variance Diagnosis Report

### Metrics
| Metric | Value |
|--------|-------|
| Training Error | 8% (100% - 92%) |
| Dev Error | 24% (100% - 76%) |
| Human Baseline | ~5% (100% - 95%) |
| Avoidable Bias | 3% (8% - 5%) |
| Variance | 16% (24% - 8%) |

### Diagnosis
**Primary Problem:** HIGH VARIANCE

**Evidence:**
- Large gap (16%) between training and dev error
- Training error (8%) is reasonably close to human baseline (5%)
- Model is fitting training data well but not generalizing

### Recommended Actions (Priority Order)
1. **Get more training data** - With 5,000 examples and 16% variance, more data is the most reliable fix
2. **Add regularization** - L2 regularization or dropout will reduce overfitting while you collect more data
3. **Data augmentation** - If getting more real data is expensive, augment existing data
4. **Reduce model complexity** - Consider fewer layers or units, though this may increase bias

### What NOT to Do
- Do NOT increase model size - you have low bias, bigger model will increase variance
- Do NOT train longer without regularization - will overfit more

### Next Steps
Start by adding dropout (0.3-0.5) and L2 regularization. Measure new training/dev error. If variance persists, prioritize data collection.

---

## Integration

This skill is part of the **andrew-ng** expert's methodology. It should be used:
- At the start of any model optimization cycle
- Whenever a user is uncertain about how to improve model performance
- Before attempting any architectural changes or hyperparameter tuning

Combine with `andrew-ng--error-analysis-protocol` when the diagnosis indicates data quality issues may be contributing to the problem.