Sales prediction using time series and XGBoost algorithm: Application of Agricultural Credit Cooperatives of Turkey

In company management, in order to make the right decision, using historical data is a successful methodology. Today, managers can easily access historical data. Thanks to the advanced enterprise programs used in the business processes of the companies, large databases that store historical records have been created. The sales process is one of the most important processes of companies. It is possible to access sales records and use these records for forecasting. Estimating the amount of product sales in retail stores, increases profitability by ensuring the effective use of company resources. Coal is a product with a high storage cost, which is sold in tons. Coal is widely used for heating in Turkey during the winter months. In addition to the cost of storage, the coal loses weight, making sellers losses in the summer. Accurate sales forecast prevents the company from making a loss. Time series methods used in statistics are widely used in sales forecasting studies. SARIMAX and Holt-Winters methods are popular time series estimation methods, preferred in studies. FB-Prophet is a new time series forecasting method developed by Facebook. FB-Prophet is used to forecast annual, weekly and daily varying time series data. Machine learning algorithms are also used in sales forecasting studies. The XGBoost algorithm is a new algorithm that is preferred in recent machine learning competitions and gives successful results. The XGBoost algorithm is also used in forecasting studies. The aim of this thesis is to find the ideal method in order to estimate coal sales made from 21 cooperatives affiliated to Sakarya Regional Association of Turkish Agricultural Credit Cooperatives. In this paper, yearly and biennial sales amount of coal are estimated in tons using total sales values between 2014-2022. SARIMAX, Holt-Winters, FB-Prophet time series methods and XGBoost machine learning algorithm are used as estimation methods. The estimation performances of the methods used in the study were compared with MA, MAPE, MSE, RMSE, R-Square performance measurement methods. Successful results have been obtained as a result of long-term prediction with XGBoost and SARIMAX methods. Using the XGBoost method with supporting parameters gave better results than other methods.