Expected error = bias squared + variance + irreducible noise.
Simple models are biased but stable; complex models are flexible but sample-dependent. Total error is minimised between the extremes.
Where finance sits. Because return data is mostly noise, the variance term dominates quickly, so the error-minimising complexity is far lower than in high-signal domains like image recognition. That is the substantive reason to prefer simple models, rather than mere conservatism.
Regularisation deliberately adds bias to remove more variance - profitable whenever the trade is favourable.