This paper presents a simulation-based procedure of forecasting abnormal mechanical actions (effects) induced during man-made accidents. The procedure is based on the so-called predictive, epistemic approach to risk assessment. It provides formal means for combining hard data and subjective information and allows forecasting the abnormal actions in the form of mathematical models, which quantify epistemic (state-of-knowledge) uncertainties in characteristics of the actions. A result of this forecasting can be used to form (quantify) an engineer's subjective confidence in future values of abnormal actions. The epistemic models allow a rough, knowledge-based estimation of probabilities of damage from abnormal actions. These models are considered to be the first step towards preventing (reducing) losses associated with damage from abnormal actions. (C) 2005 Elsevier Ltd. All rights reserved.