Forecasting the sales volumes at stainless steel sector by using data mining methods

The demand of stainless steel in our country is substantially fulfilled with importation. Sales forecasting of stainless steel is important both for the importer companies operating in the sector and for national economy because of being a widely used non-substitutional expensive and imported product. In this study, the sales forecasting of an importer company which is active in stainless steel sector and selling midproducts of stainless steel to the firms in different sectors has been examined. The data on company sales were reorganized according to sectors and aggregate. Macroeconomic factors affecting total sales and sectoral sales by model tree method were determined and the predictive models informing about the past, present and future were formed. By hierarchical clustering analysis, sector movements were examined and hierarchical time series including sales movements were obtained by grouping similar sectors. Sales forecasting of hierarchical time series were made by ARIMA and Exponential Smoothing methods. The sales forecasting of total sales and sectoral sales were modelled by Support Vector Regression (SVR), Gaussian Process Regression (GPR) and Multi Layer Perceptron (MLP) methods. 16 sectors which buys stainless steel from the corporation were determined by matching the customer codes and corporation's daily sales data at the preprocessing phase of data. Data sets containing monthly sales information for Construction, Industrial Kitchen, Metal, Machinery sectors according to the sales strategy of the corporation were prepared. The sectors which are staying out of these sectors and those which have little shares on sales were gathered under the name of Other Sectors. Data preparing phase was completed by adding the independent variables which affect the sales of the stainless steel sector. It was observed by using Model Trees Method that the most important factor affecting the sales of the Industrial Kitchen Sector was Industrial Endorsement Index, the most important variable affecting the sales of Machinery, Metal, and Construction Sectors was Producer Price Index, and the most important one affecting the Total Sales and the sales of Other Sectors was Dolar Exchange Rate. Sales figures of Construction, Industrial Kitchen, Metal, Machinery and Other Sectors were gathered in 2 sets due to similarity characteristics after Hierarchical Clustering Analysis. As being the biggest cluster Machinery, Metal and Industrial Kitchen sales occured in the first cluster (A node), and in the second cluster (B node) includes Construction and Other sector sales. Hierarchical time series prediction was made for 8 nodes consisting of 5 sectors and A node, B node and Aggregated Sales in hierarchical pattern by R software. Totally 32 models were formed by ARIMA and Exponential Smoothing methods and two combination methods as Bottom up and Optimal Combination methods. According to this just 6 of 32 models' MAPE values were calculated less than or equal to 10%. Therefore it can be said that the time series methods applied in this study were partially successful. The predictions made by SVR, MLP, GPR methods, MAPE values were calculated between the intervals of 5,53%-13,91% in the first month, 5,05%-15,57% in the second month and 5,41%-16,9% in the third month. In the prediction of first month the most succesful one for Aggregated Sales actualized by MLP method (MAPE value 5,53%). In addition, 15 of 18 models' MAPE values were calculated below 10%. The most successful prediction for the second month was for the Other Sector Sales by GPR method (MAPE=5,05%). In addition, for 13 models amongst 18, MAPE values were calculated below 10%. The most succesfull prediction for the third month prediction realized for the Other Sector Sales by GPR method (MAPE value = 5,41%). In addition, MAPE values were calculated below 10% in 13 models of 18. Having analyzed all the prediction periods it can be said that SVR was succesfull in 16, MLP was succesfull in 15 and GPR was succesfull in 10 of 18 prediction models. It could be concluded that the SVR method was more succesful than the others to forecast the Aggregated Sales and the sales of each sectors. Keywords: Data Mining, Model Trees, Clustering Analysis, Time Series, Sales Forecasting.