The shock scenarios used in the practice of stress testing are often based on ad hoc hypotheses concerning the evolution of risk factors. Consequently, they have been criticized for reflecting mild and temporary shocks maintained over very short periods. The aim of this article is to propose methodological improvements to the design and calibration of initial shocks based on the detection of outliers and structural breaks in (macro)economic and financial data. Our results of the real-world implementation of outlier detection algorithms show that the dynamics of shocks, the length of the stress horizons, and their magnitude are sensitive to the type and nature of the considered risk factor. The inferred stress horizon is longer (one to two years) for macroeconomic variables like GDP, real estate or oil prices, than for interest rate variables and the slope of the yield curve (six months).