Bias is error from a model too rigid to represent the truth, which is underfitting. Variance is error from a model so flexible that it changes a lot with the sample, which is overfitting. Total error is the sum, so there is an optimum somewhere in between.
In most machine-learning domains the optimum is quite flexible, because signal is strong and data is plentiful. In trading the signal is a rounding error on the noise, which pushes the optimum toward models with very few effective degrees-of-freedom.
Practical translation: prefer linear models, few features, and heavy regularisation. If a complex model beats a simple one on a properly purged validation set by a small margin, the simple one is usually still the better live choice because its failure modes are visible.
Related: overfitting, underfitting, regularisation, degrees-of-freedom