ANALYSIS OF MACHINE LEARNING MODELS FOR FORECASTING RETAIL RESOURCES

Abstract

The object of research is the process of forecasting loosely structured data of retail artifacts by means of machine learning. The paper analyzes data and models for forecasting retail resources. The analysis is carried out for a specific business situation and task, when a large corporation needs a fuller loading of its own warehouses with goods and resources that will be used in future periods for sale or in projects. The task is to reduce overall corporate costs by purchasing the necessary goods/resources in advance. The data required for forecasting, their sources and properties are defined. It is shown that the data will come from different sources, will have a different time interval, categorical component and logistic reference. RNN, LSTM, Random Forest, Gradient Boosting, XGBoost models and forecasting methods were chosen for such data. They were analyzed according to the criteria of data source, time interval, categorization of data, availability of a logistic component, flexibility of tools in working with heterogeneous data, requirements of tools for computing resources, interpretability of modeling results. Data sources explain where the data for analysis comes from. Usually it is: stores, warehouses, logistics companies, projects and strategic plans of the corporation. The time interval characterizes the frequency and regularity of receiving data for analysis. The criterion "data categorization" characterizes how this type of data affects the quality of the analysis. The logistic parameters of the data also characterize the impact on the analysis. "Flexibility in working with heterogeneous data" shows the ability of the model to effectively work with data of different formats and sources. Requirements for computing resources determine their necessary power for training and operation of the model. Interpretability of a model characterizes its ability to explain how and why it makes specific decisions or predictions based on input data. The more complex the model, the more difficult it is to interpret. In the retail business, interpretability is important for explaining demand forecasts. Based on the results of the analysis, the XGBoost model was recommended as the best for forecasting retail resources.