This chapter exposes deficiencies of the hypothesis-testing framework for the task of statistical verification of economic theories. The exposure shows how vital it is for model formulation to follow the machine learning approach, how diagnostic function of hypothesis testing takes precedence over its originally intended function and where the limitation of those diagnostic tests lie during the learning process. Statistical tools designed for data exploratory purposes are naturally more useful than those for confirmative purposes. Meanwhile, the necessity of learning places faithful model formulation at the core of econometric research. This is demonstrated through two examples: the first is pertinent to selection-bias concerns over the handling of incomplete cross-section samples in microeconometrics, and the second is on the appropriation formulation of leading indicator modelling issues using heterogeneous time-series samples in macroeconometrics.