This thesis explores the real-time dynamics of learning about breaks by utilising different datasets, i.e. simulated and actual (aggregate and firm-level). I am interested in the real-time identification because of its relevance for forecasting. Essentially, I raise three main empirical questions: How often do we encounter mistakes in real-time identification of breaks? How soon do mistakes get corrected in real time? How soon do we learn about the breaks in real time? I compare the effectiveness of different break models and techniques for optimal (discrete) break identification. I find that mistakes are encountered when the true breaks are not observed and when the breaks that are not the true breaks are observed in real time. By using simulated and (actual) aggregate-level datasets for the processes related to the growth rate, mistakes are encountered more often for the break model of unit root. As for the (actual) firm-level dataset of dividend series of (selected) V .S. firms, I observe that mistakes are encountered more often for the break model of trend stationary. Consistently, sequential hypothesis testing for optimal breaks are shown to make fewer mistakes compared to the information criteria used in this study. Moreover, I show that it takes several years to find the true breaks and the collection time for mistakes is usually less than a year. The learning time about the breaks and correction time for mistakes in real time are shown to be longer for the unit root model in the processes related to the growth rate for simulated and (actual) aggregate-level datasets. For the firm-level dataset, the learning and correction time are longer for the trend stationary model in the quarterly compounded process of the firm-level dividend. The learning and correction time by sequential hypothesis testing for optimal breaks are consistently shown to be shorter compared to the information criteria.