Due to the fast development and availability of computers, predictive approaches are increasingly used in the evaluation process of hazardous substances complementary to experiments. Their use was recommended as alternative to experimental testing by the REACH regulation to complete the lack of knowledge on properties for existing substances that must be registered before 2018 (upon quantities). Among the proposed predictive approaches, Quantitative Structure Property Relationships (QSPR) are powerful methods to predict macroscopic properties from the only molecular structure of substances. In that context, the HAZPRED project (2015-2018, founded by the SAF€RA consortium) aims to develop theoretical models (e.g. QSPR) and small-scale tests to predict complex physico-chemical properties (e.g. thermal stability, explosivity) of hazardous substances to complete the lack of knowledge on these hazardous substances quickly or to understand their decomposition behaviour better. In particular, this contribution will present the work done in this project on the physical hazards of organic peroxides and self-reactive substances: gathering of existing experimental data, new experimental campaigns, review of existing models and proposition of new estimation methods.