Pré-publication, Document de travail: This paper extends dynamic factor models by explicitly incorporating outliers, moving beyond conventional data screening practices. The methodological contribution includes introducing fat tails and outliers multiplicatively into innovation volatility, and two distinct approaches for modelling outliers are presented to address large jumps. Empirical findings demonstrate that outlier-augmented models consistently outperform benchmark models in point and density forecasting, with the most significant improvements observed in nowcasting horizons. Incorporating outliers becomes particularly crucial during major crises, enhancing forecasting accuracy by 44% compared to the benchmark. The uniform-mixture approach is found to be more robust than the student-t models, as it targets extreme variations without disrupting the smoothness of the stochastic volatility process.