Bias-variance tradeoff: practical rules of thumb

Bias is error from wrong assumptions (underfitting); variance is sensitivity to training-data fluctuations (overfitting). Total expected error = bias^2 + variance + irreducible noise. Practical rules of thumb: train-eval gap <1% = high bias; >5% = high variance. If high bias: bigger model, better features, more layers, longer training, less regularization. If high variance: more data, stronger regularization, dropout, shorter training, ensemble (bagging), data augmentation. Senior nuance on real production models: deep nets without enough data exhibit high variance; classical models (linear, GBMs) exhibit high bias. Quick diagnostic: train on 10% of data; if train error spikes, the model is probably data-limited (variance); if it stays low, you have capacity but bad features (bias). Ensemble methods like bagging reduce variance; boosting reduces bias.