Something maybe obvious but worth repeating is that there are 2 kinds of errors: predictable and unpredictable. Bias is the predictable error, it's the direct

Something maybe obvious but worth repeating is that there are 2 kinds of errors: predictable and unpredictable. Bias is the predictable error, it's the direction we are likely to be wrong in. In many practical applications unbiased error is not what we want. If the cost of being wrong in different directions is asymmetric then we want to be biased so that our mistakes are less costly. The unpredictable error is noise. In this example we are trying to create something maximally unpredictable, so the goal is to remove all biases, giving pure noise.