Box-Cox transformation is a statistical technique that transforms your target variable so that your data closely resembles a normal distribution.
In many statistical techniques, we assume that the errors are normally distributed. This assumption allows us to construct confidence intervals and conduct hypothesis tests. By transforming your target variable, we can hopefully normalize our errors, if they are not already normal.
Additionally, transforming our variables can improve the predictive power of our models because transformations can cut away white noise.
What Is Box-Cox Transformation and Target Variable?
Suppose we had a beta distribution, where alpha equals one and beta equals three. If we plot this distribution, then it might look something like below:
plt.figure(figsize = (8, 8))
data = np.random.beta(1, 3, 5000)
sns.distplot(data)
plt.show()
We can use the Box-Cox transformation to turn the above into as close to a normal distribution as the Box-Cox transformation permits.
tdata = boxcox(data)[0]
plt.figure(figsize = (8, 8))
sns.distplot(tdata)
plt.show()
And now our data looks more like a normal distribution.
What Is a Target Variable?
A target variable is the variable that you are trying to estimate. What’s nice about the linear regression model is that your target variable can take on virtually any form: continuous, ordinal, binary and more.
However, different forms of target variables will yield different interpretations of your slope parameters. For example, if your target variable is binary — that is, it takes on a value of either one or zero — then the slope parameters of your regression model represent the way a one-unit increase in your independent variables changes the probability that your target variable will equal one.
What Is the Box-Cox Transformation Equation?
If w is our transformed variable and “y” is our target variable, then the Box-Cox transformation equation looks like this:
In this equation, “t” is the time period and lambda is the parameter that we choose. You can also perform the Box-Cox transformation on non-time series data.
Notice what happens when lambda equals one. In that case, our data shifts down, but the shape of the data does not change. If the optimal value for lambda is one, then the data is already normally distributed, and the Box-Cox transformation is unnecessary.
How Do You Choose Lambda?
We choose the value of lambda that provides the best approximation for the normal distribution of our response variable.
SciPy has a boxcox function that will choose the optimal value of lambda for us.
scipy.stats.boxcox()
Simply pass a 1-D array into the function and it will return the Box-Cox transformed array and the optimal value for lambda. You can also specify a number, alpha, which calculates the confidence interval for that value. For example, alpha = 0.05 gives the 95 percent confidence interval.
If llf is the log-likelihood function, then the confidence interval for lambda can be written as:
In this equation, “X²” is the chi-squared distribution. It may also be unnecessary to transform your data if the confidence interval includes one.
Next, fit your model to the Box-Cox transformed data. You must revert your data to its original scale when you are ready to make predictions.
For example, your model might predict that the Box-Cox transformed value, given other features, is 1.5. You need to take that 1.5 and revert it to its original scale — the scale of your target variable.
Thankfully, SciPy also has a function for this.
scipy.special.inv_boxcox(y, lambda)Enter the data you want to transform, “y,” and the lambda with which you had transformed your data.
Limits of Box-Cox Transformation
If interpretation is your goal, then the Box-Cox transformation may be a poor choice. If lambda is a non-zero number, then the transformed target variable may be more difficult to interpret than if we simply applied a log transform.
A second issue is that the Box-Cox transformation usually gives the median of the forecast distribution when we revert the transformed data to its original scale. Occasionally, we want the mean, not the median, and there are other ways to do that.
Now you know about the Box-Cox transformation, its implementation in Python, as well as its limitations.
