Sales Forecasting

The customer needed a reliable and store-specific forecasting solution for its chain of 16 physical shops located in different cities across Ukraine. Accurate forecasts were required to support operational planning at each location. The focus was exclusively on the offline stores, with the aim of improving inventory control, optimizing storage usage, and enhancing overall resource management. The customer’s existing forecasting methods were insufficient for addressing variations in local demand, seasonality, and external factors.

A machine learning-based forecasting system was developed to meet these needs. Historical sales data from the financial database was enriched with external inputs such as weather information and holiday calendars.

Multiple modeling approaches were tested and evaluated:

• Transformer-based models (e.g., TimeSformer, Autoformer)
• Recurrent Neural Networks (LSTM, GRU)
• Classical time series models (ARIMA, Prophet)
• Tree-based models and ensembles (XGBoost, LightGBM, Random Forest)

The implemented solution provided detailed and reliable daily and monthly sales forecasts for each of the customer’s 16 physical shops. By leveraging historical sales data alongside external factors such as holidays, weather conditions, and marketing activities, the forecasting model enabled a more data-driven approach to store-level operations.

As a result, the customer was able to significantly improve inventory planning by aligning stock levels with expected demand, thus reducing the risk of overstocking or shortages. The forecasts also supported better utilization of storage space, allowing for more efficient logistics and supply chain coordination. Moreover, the integration of the forecasting outputs into the customer’s existing Power BI environment gave store and inventory managers clear visibility into upcoming demand trends, facilitating timely and informed decisions.