Comparison of machine learning based demand forecasting models

In order to prevent delays in production plans, keeping the raw materials and semi-finished products in the specified amount has a significant impact on the realization of the plans. In this regard, the importance of demand forecasting models in ensuring continuity in the supply chain and achieving customer delivery dates cannot be denied. Nowadays, as developments occur very rapidly and it becomes more and more difficult to keep up with developments, it has become a greater need for businesses to predict the future in order to maintain their existence. Being able to react to demands in a timely manner by adapting to changes in customer volumes and developments in technology contributes to the sustainability of the supply chain. It is possible for businesses to secure their future even in a highly competitive environment with a successful supply chain management. High variability in demand may cause disruptions in supply and inventory management. These disruptions may lead to increased costs due to excess stock or order losses as a result of products not being delivered. As a result of failure to control the disruptions, the situation may turn into customer loss for the business. Demand forecasting is one of the most needed functions in solving these problems. Demand forecasting methods offer advantages in many areas such as planning, cost and inventory management. With these advantages, companies' inadequacies decrease and their resilience in the market increases. The development of realistic demand forecast models, which is one of the necessary studies to prevent order delays, has an important place in the literature and industrial reports. In this study, it is aimed to develop a realistic demand forecasting model based on machine learning. With this study, determining the primary raw materials and semi-finished products at the required time and quantity in order delays needed for products will significantly reduce the problems. Despite the complex problems involved in demand forecasting processes, the common goal in demand forecasting is to benefit businesses by making a future forecast with high accuracy. At the same time, thanks to accurate forecasts, the efficiency of businesses increases. More accurate determination of product quantity and variety places demand forecasting at the center of planning. Compared to traditional forecasting methods, machine learning-based forecasting methods provide better results. In this study, which was conducted using real sales data through a case study, solutions were realized with selected machine learning-based methods. It is aimed to determine the method that gives the optimum result with the established models. The data to be used in demand forecast models were selected among the parameters that directly or indirectly affect the demands of products whose main component is steel. Euro/dollar parity was chosen as one of the independent variables, because the raw materials used in the production of the products are traded internationally in dollars and sales are made in euros. Since approximately 80% of the cost of the products comes from raw materials and the main component of the products is steel, the price of steel has become another independent variable. Energy cost was used as another independent variable, as the most important cost item in the cost of the products after the raw material and machinery investment cost. Additionally, there was a need to add recent data to increase the accuracy of the demand forecast model. Thus, in the models, actual order data for each product one, two and three months ago was also evaluated among the independent variables. The actual sales data used in the demand forecast model was taken from a large-scale international company selling products in the steel industry. In addition to the data of the last three periods, the attribute was created with the moving average data of the last three months. In the re-model simulations, it was observed that the addition of these features significantly contributed to the reduction of errors. In addition, it was realized from the simulations that the change in variables should be included among the attributes in order for the model to follow the changes in orders. In this regard, the differences in steel prices between periods, the differences in energy costs between periods and the change values of orders between periods were added to the independent variables. When deciding to use independent variables in the models, various modeling simulations were made with the available data. Comparisons of artificial neural networks, decision forest and linear regression methods were made to evaluate the performance of the demand forecasting model. Statistical methods have been applied to the independent components used in artificial neural networks, decision forest and linear regression models to obtain more accurate results in the models. In the training set, changes such as replacing the distribution of the data that disrupted the series and were at the extreme ends of the distribution with the average value, normalization, and logarithmic transformations were made. Outlier data in the data set used in the model were replaced with average values. While the values at the extremes were removed, the values above 95% and below 5% were selected. In addition, since the range of real sales values used in training the model was very wide, logarithmic transformation was performed to eliminate the skewed distribution in the data and obtain a distribution close to the normal distribution. In order to determine the most relevant features in the development of demand forecast models, 6 different feature selection methods were applied: Particle Swarm Optimization, Harris Hawk Optimization, Grey Wolf Optimization, Dragonfly Optimization, Genetic Optimization and Gravity Search Optimization. By making feature selections with these optimization algorithms, models that give more accurate results have been developed. In the developed models, not enough raw materials and semi-finished products may cause order delays, while their excess may cause an increase in stock costs. In the developed model, it is predicted that order delays and inventory costs can be reduced by reducing forecast errors. As a result of this study, it will be ensured that the stock quantities of the materials are kept at the optimum level. Based on the results of this study, it can be said that overorders and delays in product delivery can be significantly reduced. Considering all the findings in the literature, in this study, experiments were conducted with data splitting rates of 70-30%, 80-20% and 90-10% in training the models. Test studies have shown that the best results are achieved with the 90-10% option. It is predicted that the reason for this difference from the 80-20% rate found in the majority of the literature is that the number of periods of available data is relatively less. 89 months of data were collected, 80 months of which were used to train the model, and 9 months of data were used as test and performance data. The models developed within the scope of this study were compared according to different demand forecasting methods. In this study, 84 (4x3x7) different models were developed for 4 different products selected for welded steel manufacturing products, 3 different machine learning algorithms and 7 cases with and without feature selection, and these models were evaluated in terms of performance criteria. When the developed models were evaluated according to MAPE, MAE, RMSE and MSE performance criteria, it was evaluated that the model developed with the artificial neural networks method gave better results than other models. Four models built with the artificial neural networks method made predictions with high accuracy. With the decision forest method, highly accurate forecast was obtained in 3 models and good forecast was obtained in 1 model. Among the models created with the linear regression method, highly accurate forecast was obtained in 1 model and good forecast values were obtained in 3 models. It was observed that the model developed with artificial neural networks reached an average accuracy of 91.9%, the model developed with the decision forest method reached 91.3%, and the model developed with the linear regression method reached 87.3% accuracy. The artificial neural networks method predicted the highest and lowest demands better than other methods. It is predicted that if the artificial neural networks method is used in other demand forecasting studies, as in this study, it will achieve the objectives of the research by providing highly accurate forecasts. If the decision forest method is used, generally highly accurate forecasts can be obtained, as in artificial neural networks. It can be said that the the forecasts obtained with decision forests will similarly provide the desired benefits to the business at a slightly lower level than the artificial neural networks method. It can be predicted that the linear regression model can provide good forecasts, but when error rates are encountered, it does not need to be preferred over other methods. As a result, it has been observed that the artificial neural network model is a more accurate choice for the desired contributions.