The ability to accurately model sales time series plays an important role in the analysis routines of e-commerce businesses and has numerous applications in areas such as sales forecasting or in the detection of non-trivial relations between products. However, the stochastic nature, presence of artefacts from real world events (e.g. large sales events or server crashes) and the great range in purchase density caused by popularity differences make the analysis of sales data challenging, and flexible yet robust modelling procedures are required. In this thesis, we address these challenges in an application driven setting to replace a heuristic method for sales time series modelling in an item recommendation system developed by the data analysis company Sift Lab. We propose a novel piecewise constant regression (PCR) method to estimate expected purchase rates in sales time series using the Minimum Description Length (MDL) Principle for model selection. The PCR method searches the space of piecewise constant Poisson and Zero-Inflated Poisson (ZIP) models for an increasingly good approximation of the underlying structure of the data and uses a binary segmentation algorithm to navigate the space in an efficient manner. As selection criteria, we derived and evaluated three different coding schemes based on MDL theory. We evaluated the method in two settings: First, we compared the coding schemes on synthetic data and analysed the computational complexity of the binary segmentation algorithm. Second, we compared the PCR and the heuristic method on real sales time series data. An out-sample quality of fit test was performed, and the respective impact of each method on the recommendation system pipeline was assessed. Overall, the PCR method captured apparent structures in the data well but struggled with gradual behavioural changes and generally tended to underfit. A stochastic complexity encoding was found to provide the best model selection approach, supported by consistent results on both synthetic and out-sample quality of fit tests. Furthermore, the results suggested that Poisson noise was a more suitable modelling assumption than ZIP noise for the dataset considered. We found comparable performances for the PCR and heuristic method, both in terms of quality of fit and on recommendation system performance. However, the relative performance of the respective MDL codes did not align well between the two tests, suggesting that estimate and recommendation quality are not as directly connected as originally expected.