Underfitting and OverfittingConclusion
Building a good model is a balancing act. Too simple, and it ignores the pattern: high bias, underfitting. Too flexible, and it memorises the luck of its training data: high variance, overfitting. Your job is to turn the dials — the degree of a curve, the span of LOESS, the of KNN — to the point where the two balance, and to judge that point on data the model has not seen.
There is one more way to tame a very flexible model: add a penalty that keeps its numbers small, so it cannot bend too much. This is called
Speed Round: Underfit or Overfit?
Get five right to move on.