juniorModel Evaluation

What is a baseline model?

Updated May 17, 2026

Short answer

A baseline model is a simple model or rule-based approach used as a reference point for evaluating the performance of a more complex machine learning model. It establishes the minimum performance level that a new model should beat to be considered useful. A baseline helps determine whether the added complexity of a machine learning solution actually provides meaningful improvement.

Deep explanation

A baseline model is the first model or strategy created before building a more advanced machine learning system. It provides a benchmark that answers the question:

"Is our machine learning model actually better than a simple approach?"

Without a baseline, it is difficult to know whether a complex model is providing value. A model achieving 85% accuracy may seem good, but if a simple baseline achieves 90%, the model is not useful.

Why baseline models are important

Baseline models help with:

  • Measuring improvement
  • They provide a starting point for comparing new models.
  • Improvements can be quantified instead of judged subjectively.
  • Detecting unnecessary complexity
  • A complicated model may require more data, computing resources, and maintenance.
  • If it does not significantly outperform a simple baseline, the extra complexity may not be justified.
  • Identifying problems early
  • If a sophisticated model performs worse than a baseline, it may indicate issues with data quality, feature engineering, or model selection.
  • Setting realistic expectations
  • A baseline shows how difficult the problem is.
  • Some tasks are hard because even simple approaches perform poorly.

Types of baseline models

The appropriate baseline depends on the problem.

1. Random baseline

A random baseline makes predictions randomly.

Example:

  • For a binary classification problem, randomly guessing between positive and negative classes.

This is useful as a minimum reference point but is rarely a practical solution.

2. Majority class baseline

For classification problems, the model always predicts the most common class.

Example:

Dataset:

TEXT
90%: Not fraud
10%: Fraud

Baseline prediction:

TEXT
Always predict: Not fraud

This model achieves 90% accuracy, but it completely fails to detect fraud cases.

3. Simple rule-based baseline

A baseline can use manually designed rules.

Example:

Python
if account_age < 7_days:
    predict = "fraud"
else:
    predict = "not fraud"

This may be simple but can provide a useful comparison point.

4. Simple statistical model

A basic machine learning model can also serve as a baseline.

Examples:

  • Linear regression for a regression problem.
  • Logistic regression for classification.
  • Decision tree with limited depth.

A more advanced model, such as a neural network, should demonstrate meaningful improvement over this simpler approach.

Evaluating against a baseline

The baseline should be evaluated using the same data splits and metrics as the final model.

Example:

TEXT
Baseline model:
Accuracy: 82%


New machine learning model:
Accuracy: 88%


Improvement:
+6 percentage points

The improvement should also be considered in terms of cost and complexity.

A small improvement may not be worthwhile if the new model requires significantly more infrastructure or maintenance.

Baseline models and data leakage

A baseline should be created carefully to avoid unrealistic comparisons.

For example:

  • The baseline should not use information unavailable at prediction time.
  • It should be evaluated on the same test data as the final model.
  • It should follow the same problem definition.

Baseline versus benchmark

These terms are related but slightly different:

  • Baseline
  • A simple starting point created for comparison.
  • Usually built by the team working on the problem.
  • Benchmark
  • A standard performance reference, often from published research or industry results.

A benchmark may represent state-of-the-art performance, while a baseline represents the minimum acceptable starting point.

Real-world example

Suppose an online store wants to predict whether a customer will buy a product after viewing it.

A simple baseline could be:

Python
def baseline_model(previous_purchase_rate):
    return previous_purchase_rate > 0.5

If historically 30% of visitors purchase a product, the baseline might predict:

TEXT
Every visitor will not purchase.

Performance:

TEXT
Baseline accuracy: 70%

A machine learning model using customer behavior, browsing history, and product information achieves:

TEXT
Machine learning model accuracy: 82%

The baseline shows that the machine learning model improved performance by 12 percentage points. The company can then decide whether this improvement justifies the cost of deploying and maintaining the model.

Common mistakes

  • * Comparing a new model without first establishing a baseline.
  • * Using an overly strong baseline that already uses information unavailable during prediction.
  • * Using only accuracy for a baseline when the dataset is highly imbalanced.
  • * Assuming a complex model is automatically better than a simple baseline.
  • * Creating a baseline that is evaluated on different data than the final model.
  • * Ignoring operational costs when deciding whether improvement over the baseline is valuable.
  • * Treating a baseline as the final production solution instead of a comparison point.

Follow-up questions

  • Why is a baseline model necessary when building a machine learning system?
  • What makes a good baseline model?
  • Why can a majority class classifier be a misleading baseline?
  • Should a baseline model always be a machine learning model?
  • How do you know if a new model is better than the baseline?
  • Can a baseline model ever be used in production?

More Model Evaluation interview questions

View all →