Development of Predictive Models to Optimize Order Forecasting in a Technology Company

This thesis focuses on the development of a predictive model for orders by applying methods for time series analysis. The objective was to automatically preprocess data for data- differentiated customer groups and find a model which could appropriately capture differences among customers. Development of the predictive model was based on the CRISP-DM methodology, which provided structure from business understanding to deployment. Within the theoretical section, the individual approaches for time series prediction were presented and later applied in the practical section. The key step for the development of the predictive model was preprocessing of data. As part of the preprocessing, data were logarithmically transformed, exponentially smoothed, smoothed using moving averages, and then prepared with pipelines. Since the data exhibited significant fluctuations, the best results were obtained with exponential smoothing using wider windows. The algorithm XGBRegressor was selected based on the minimum error on the validation data and was fine-tuned with Grid search, due to time constraints and the parallel processing capabilities it provides. Based on the results, the key steps were data aggregation at the monthly level and preprocessing of the target and features. The XGBRegressor model is currently being tested to verify its stability and reliability in a real environment.