Bias metric is a descriptive statistics measure that is computed before and after training the model. There are two extremes of the bias metric. A lower bias score denotes that the model assumptions or predictions are almost correct. Another conclusion that is drawn is that the predictions are equally split into two halves, with one above the actual value, while the other half is below. However, if the model throws a higher bias score with low variance, then the architecture needs to be revisited.