> ## Documentation Index > Fetch the complete documentation index at: https://notes.kodekloud.com/llms.txt > Use this file to discover all available pages before exploring further. # Creating Performance Baselines > Guide on creating and using baseline models to evaluate machine learning performance, from naive and heuristic to model-based and SOTA baselines, with examples and a practical checklist. A baseline model is a simple, well-defined reference used to evaluate whether a new model provides meaningful improvement. In practice, you feed the same input to both a BaselineModel and your TrainModel and compare their predictions and metrics. The purpose of a baseline is to establish a minimum acceptable performance and to justify added model complexity when the trained model exceeds that baseline. The image illustrates a flowchart showing a baseline concept overview, where input data is processed through a baseline model and a trained model to generate and compare predictions. ## Common baseline types Use increasingly stronger baselines as you progress: start simple and advance toward more competitive references. Below is a concise description of common baseline classes. | Baseline Type | Purpose / When to Use | Typical Example | | ----------------- | --------------------------------------------------: | -------------------------------------------------- | | Naive | Quick sanity check to set a very low bar | Always predict the majority class (classification) | | Random | Establish a minimum floor for performance | Random class labels or random regression values | | Heuristic | Capture simple domain rules; interpretable | Rule-based formula like `y = 2x` | | SimpleModel | Lightweight ML models providing data-driven fit | Linear/logistic regression, decision stump, k-NN | | HumanLevel | Compare model to human expert performance | Expert label accuracy in medical imaging | | SOTA | Benchmark against best published or industry models | Top-performing research model or leaderboard entry | | Business-as-usual | Compare to the system currently in production | Existing model or rule set serving customers | Choose baselines appropriate to your problem and dataset. Start with the simplest (naive) and progress to heuristics, simple models, and SOTA baselines until you establish a realistic target for development and deployment. ## Why baselines matter (example) Suppose a baseline reaches 70% accuracy and your trained model reaches 85% accuracy — a 15 percentage-point improvement. That gap quantifies the model’s added value and helps decide whether extra complexity is justified. The image illustrates a comparison between a baseline model with 70% accuracy and a trained model with 85% accuracy, highlighting a 15% improvement in accuracy indicating meaningful learning. ## Naive baselines by problem type Naive baselines are fast to compute and useful as the first checkpoint. They intentionally ignore patterns in the features and act as a simple reference. | Problem Type | Naive Strategy | | ------------------------ | --------------------------------------------------- | | Classification | Predict the majority class | | Regression | Predict the global mean or median of the target | | Time series | Predict the last observed value (persistence model) | | Ranking / Recommendation | Rank items by global popularity | The plot below illustrates a naive regression baseline predicting the global mean as a constant (dashed line) while the actual target values vary over time. This highlights how naive approaches provide a minimal benchmark that more sophisticated models should surpass. The image is a line graph comparing "Naive Regression Baseline vs Actual" values over different time steps, with actual values depicted as a solid line and the naive baseline as a dashed line. The graph includes labeled axes for target value and time step. ## Heuristic baselines Heuristic baselines apply simple, interpretable rules derived from domain knowledge (for example, `y = 2x`). They often outperform naive baselines because they encode known relationships and provide a stronger, explainable benchmark. Use these to confirm whether model complexity is adding value beyond simple rules. ## Model-based baselines Model-based baselines use basic machine learning algorithms to create a data-adaptive reference. These typically outperform naive and heuristic baselines and are essential for structured data problems. Common simple-model baselines: * Linear regression / logistic regression for tabular data * Decision stumps (single-level decision trees) * k-Nearest Neighbors (kNN), often with k=1 or small k These model families are lightweight, fast to train, and help set realistic expectations before moving to more complex architectures. The image describes three simple machine learning models: linear/logistic regression for structured data, decision stumps (single-level decision trees), and nearest neighbor approaches (kNN with k=1). A model-based baseline such as linear regression adapts to the data and typically produces a better fit than a heuristic rule. The example below shows a linear baseline described by `y = 2.00x + 0.43` (dashed line) compared to actual values — a clear, data-driven reference. The image shows a graph comparing model-based baseline values (a dashed blue line) and actual values (an orange line with dots) plotted against a feature (x-axis) and target value (y-axis). The baseline model follows the equation (y = 2.00x + 0.43). ## State-of-the-art (SOTA) baselines SOTA baselines benchmark your approach against the best published or industry-standard models. These are useful to determine whether your model is competitive: * If SOTA = 90% and your model = 85% → further improvements needed. * If your model approaches SOTA, you may focus on deployment, inference optimization, or cost/performance trade-offs. The image compares a baseline model with 90% accuracy to a trained model with 85% accuracy, highlighting the concept of state-of-the-art (SOTA) baselines. ## Practical checklist for baselines * Establish at least one simple baseline (naive or random) before model tuning. * Iterate: naive → heuristic → simple model → SOTA as needed. * Use the baseline gap (performance delta) to quantify business value and justify complexity. * Report baselines alongside model metrics in documentation and model cards for transparency. * Re-evaluate baselines when data, labels, or business objectives change. Resources and further reading: * [Model evaluation best practices](https://en.wikipedia.org/wiki/Cross-validation_\(statistics\)) * [When to compare with SOTA models](https://www.semanticscholar.org/) By defining clear baselines early, you save development time, avoid unnecessary complexity, and focus on changes that deliver measurable value.