What is the 'bias-variance tradeoff' in machine learning?

Bias is the systematic error from incorrect assumptions. High bias models (linear regression on non-linear data) underfit. Variance means sensitivity to training data fluctuations. High variance models (deep decision trees) overfit.  Total Error = Bias² + Variance + Irreducible Noise.  As model complexity increases Bias decreases and Variance increases.  Optimal model complexity minimizes total error.  Solutions to this problem are Regularization (reduces variance), Ensemble methods (Bagging reduces variance, Boosting reduces bias) and More data (reduces variance). In banking, a credit model with high bias may systematically underestimate risk for a demographic group.