Machine learning-supported cement price forecasting in the construction chemicals industry

This thesis aims to forecast the procurement prices of cement—a key raw material—for the Eskişehir factory of a company operating in the field of construction chemicals. Developed in line with consultations held with company representatives, the study seeks to provide a data-driven contribution to strategic and operational decision-making processes. Among the five fundamental raw materials used in the production of construction chemicals—namely cement, cellulose, calcite, kraft paper, and polymers—this research focuses on cement within the scope of a pilot application, specifically targeting the Eskişehir facility. The analysis is based on a monthly dataset compiled for the period between January 2013 and December 2024. A total of 30 candidate features were evaluated across different thematic categories such as supply-demand balance, energy and transportation costs, economic indicators, and demographic factors. In the initial phase, superficial missing value imputation was conducted to enable feature selection. Subsequently, two distinct feature selection strategies were adopted: one based on expert opinions (Study 1) and the other on academic and statistical methodologies supported by machine learning techniques (Study 2). Advanced missing value imputation techniques were then applied to the resulting datasets. The study was further diversified through additional modeling scenarios, including feature set merging (Study 3), the exclusion of the capacity utilization variable from the model (Study 4), and correlation-based feature reduction (Study 5). Eight different machine learning algorithms were tested during the modeling process, using both default and optimized hyperparameters. The models were evaluated using various performance metrics. The results reveal that feature selection based on domain knowledge yields particularly high prediction accuracy in machine learning models. The modeling framework developed within the scope of this thesis is recommended to be extended to other raw materials beyond cement, thereby enabling the construction of a more comprehensive decision support system.