# Always Beat a Baseline — Machine Learning Basics

Source: https://www.skillbyai.com/en/machine-learning/r-baseline

> A model is only useful compared with something simple.

## Dummy models and simple rules

Before celebrating a metric, compare it with a **baseline**: predicting the mean (regression), the most frequent class (classification), last week's value (forecasting), or the current business rule. scikit-learn provides `DummyRegressor` and `DummyClassifier`. If your model barely beats the baseline, it may not be worth deploying; if it beats it massively, check for leakage. The baseline also gives non-specialists a reference point: 20% lower error than the current rule is a clear message.

## Baselines by problem type

Pick one before modelling.

```text
regression       DummyRegressor(strategy="mean")      or the current business estimate
classification   DummyClassifier(strategy="most_frequent")
forecasting      "same as last period" / seasonal naive
ranking          popularity order
any problem      the existing rule the business uses today
```

## Report the lift

State results as improvement over the baseline, for example MAE 45 versus 58 for the mean.

**Quiz:** Why compare a model with a baseline?

- [ ] It trains the model faster
- [ ] Baselines are always better
- [x] To know whether it actually adds value over something simple
- [ ] It removes the need for a test set

*Answer:* To know whether it actually adds value over something simple. Context makes metrics meaningful.