Why does performance by bureau score change so radially through the credit cycle? Why do we have cut-off scores? Why do machine learning models degrade so fast when deployed, and do they need to? What is the real purpose of alternate data? What are the true dimensions of borrower behavior that we need to understand? Why isn't forecast uncertainty included in underwriting? Why do applications of Modern Portfolio Theory fail for loan portfolios? These questions and many more are answered in this integrated approach to credit risk analytics. Credit risk analysts are not tapping the real power of machine learning and alternate data, because their models are built in a 1960s scoring architecture. Changing the architecture not only solves problems of overfitting and out-of-time degradation, but it also turns machine learning models into cash flow forecasters that integrate directly with yield and NPV models in finance. When account-level forecasts directly predict yield, underwriting decisions can be based on financial targets directly, rather than judgmental, backward-looking score cutoffs. The material is presented conceptually with a focus on analytic methods with business value. To solve these decades-long mysteries, the industry must break free of the 1960s mindset of how analytics should be used in credit risk, and this book lights the way.