As long as you have collected data on outcomes that contains more than 100 defaults, with covariates that would allow an informed expert to distinguish between higher and lower risk transactions, it's often possible to be more accurate than an informed expert on this problem.
These Kaggle kernels contain interesting feature engineering and tuning strategies https://www.kaggle.com/c/home-credit-default-risk.
This is nice if you have a single feature that rank-orders well and you would like to calibrate https://www.chrisstucchio.com/blog/2020/isotonic_python_pack...
As long as you have collected data on outcomes that contains more than 100 defaults, with covariates that would allow an informed expert to distinguish between higher and lower risk transactions, it's often possible to be more accurate than an informed expert on this problem.