Regression

Box Cox Transformation

The box cox transformation is a parametrized transformation that tries to get distributions “as close to a normal distribution as possible” It is defined as:

$$\text{boxcox}(y_i) = \frac{y_i^{\lambda} - 1}{\lambda}$$

You can think of it as a generalization of the square root function: the square root function uses the exponent of 0.5, but box cox lets its exponent vary so it can find the best one.