Cam sektöründe talep tahmin yöntemlerinin uygulanması ve değerlendirilmesi

Abstract

Şirketler varlıklarını sürdürmek, kâr etmek ve büyüyebilmek için gelecekle ilgili sürekli araştırmalar yaparak ürünlerine olan talebi tahmin etmeye çalışırlar. Tahmin, geçmiş verilerden ve deneyimlerden yola çıkarak gelecekle ilgili öngörüde bulunabilmektir. Tahmin etme salt şirketler için değil ülkelerin ekonomilerini ayakta tutmaları açısından da oldukça önem taşımaktadır. Talep tahmini, gelecek dönemlerde tüketicilerin talep edecekleri ürün ya da hizmetin miktarını önceden kestirebilmektir. Üretim plânlama faaliyetlerinin başlangıç noktası olan talep tahmini sâyesinde şirketler, kapasiteyi optimum düzeyde kullanma, stok miktarlarını koruma, ürün bulunabilirliğini arttırarak yeğlenen şirket hâline gelme, müşteri hoşnutluğunu arttırma gibi birçok olanağı elde edebilmektedirler. Doğru tahmin girdileri ile üretim sistemleri plânlanır ve gerekli hammadde, malzeme, yarı ürün, makina-donanım, insan kaynağı ve finansman konularında kararlar verilir.Literatüre baktığımızda talep tahmini konusunda birçok yöntem geliştirilmiştir. Bu yöntemlerin doğruluk oranları; kullanılan verilerin yapısına, doğrusallık ya da karmaşıklığına, kapsadığı zaman aralığına göre değişim göstermektedir. Tahmin ediciler, geçmiş verinin oluştuğu koşullarda genellikle zaman serisi tahmin yöntemlerini yeğlemişlerdir. Bunun dışında son zamanlarda karmaşık problemlerde belirgin şekilde kullanılmaya başlayan yapay zekâ sistemlerinden yapay sinir ağı, simülasyon, genetik algoritma, bulanık mantık, gri teori, destek vektör makinaları gibi sezgisel yöntemler dikkate alınmaktadır. Amaç eldeki veri kümesine doğru tahmin yöntemini uygulayarak optimum sonuç elde edebilmektir.Bu çalışmanın amacı; ülke ekonomisine büyük ölçüde katkı sağlayan cam sektöründeki ürünlere olan talebi önümüzdeki dönem için doğru tahmin edebilmektir. Çalışmada öncelikle tahmin ve talep tahmini kavramlarından sözedilerek konuya giriş yapılmış, literatürde ele alınan yöntemler ve gelişimleri ile konu sürdürülmüştür. Sonrasında, uygulamada ele alınan sektör ve şirket tanıtılarak tahmin edilecek ürünlerin analizi yapılmıştır. Tahmin için ürün gruplarının seçimi ile ürünlerin hangi düzeyde tahmin edileceği önemlidir. Satış bölümündeki ekipler tarafından belirlenen talep tahmin birimlerine (DFU) göre ABC-XYZ analizi yapılarak, tahmin edilecek ürünlere karar verilmiştir. Tahmin yöntemlerini uygulamak için yurtiçi pazarda en büyük üretici olan cam şirketinin en fazla satış yoğunluğuna sahip olan bölgesi belirlenerek, müşteri tarafından en çok talep edilen (satılan) ürün ile ikincil işlem içeren katma değerli bir diğer ürün dikkate alınmıştır. Yöntemler uygulanmadan önce veri kümesi temizlik işlemine tâbi tutularak kampanya, fiyat zammı, indirimi, grev süreçleri gibi etkileyici etmenlerden arındırılmıştır. Uygulamada yöntem olarak zaman serisi analizlerinden hareketli ortalama, üstel düzeltme, Holt üstel düzeltme, Winter üstel düzeltme ve ARIMA yöntemleri ile, yapay zekâ ve sezgisel algoritmalardan yapay sinir ağları kullanılmıştır. Zaman serisi yöntemleri için Minitab 17 istatistik programından destek alınırken, yapay sinir ağları için Neurosolutions 5.05 Developer Excel ara yüzüyle birlikte çalışılmıştır. Her bir yöntem için kendi içinde belirlenen model tasarımlarıyla duyarlılık analizleri yapılarak hata oranlarına göre en uygun parametreler seçilmiştir. Doğruluk oranı yüksek parametreler ışığında belirlenen her bir yöntem ile iki ürün için 2015 yılı 12 aylık satış verisi tahmin edilmiştir. Literatürde kullanılan tahmin hata ölçümlerinden en yaygın olanı, ortalama mutlak yüzde hatadır (MAPE). Tahmin performansında MAPE hata ölçümünün yeğlenmesinin nedeni, farklı birimlerdeki modellerin karşılaştırılmasında yüzde ölçüm ile ortak bir nokta oluşturarak ortaya çıkabilecek dezavantajları yok etmesidir. Çalışmanın son aşamasında, ele alınan tüm yöntemlerin en yakın dönemdeki tahmin performansları, ortalama mutlak yüzde hata ölçümüne göre değerlendirilmiş ve geleceği tahmin etmek için optimum sonucu veren yöntem ortaya çıkarılmıştır.Temizlenen veriler üzerine yapılan tahmin sonrasında, konusunda uzman, deneyimli kişilerle bir araya gelerek tahmin sonuçları zenginleştirilebilir. Bu sâyede kişilerin öngördükleri farklı sonuçlar ya da birtakım ekonomik göstergeler, kampanya değişimleri gibi dış etmenler varsa bunlar yansıtılarak, satış ya da talep tahminleri son hâlini alır.

Companies try to forecast the demand for their products by means of conducting constant researches relating to future to continue their existence, to receive profits, and to grow. Forecasting is the ability to predict the future from the past data and experiences. Almost all sectors, such as health, tourism, textiles, construction, and finance, need forecasting to plan their future. Forecasting future provides the companies with numerous advantages like taking of managerial decisions, increase of productivity, evaluation of their performance, contribution to the accurate functioning of the processes, and bringing the investment requirements of them to the surface. Forecasting is important not only for the companies but also for the countries to sustain their economies.Demand forecasting is the ability to predict beforehand the amount of the products and services that the consumers might demand in the future. Thanks to demand forecasting that is the starting point of the production planning activities, companies are able to acquire many opportunities like the utilization of capacity and the stock quantities at an optimum level, being a preferred company by enhancing the availability of the products, and increasing the customer satisfaction. Through accurate forecasting inputs, production systems are planned and decisions are made with regard to the required raw material, material, semi-products, machinery-equipment, human resource, and financing.It appears in the literature that numerous methods concerning demand forecasting have been developed. Accuracy ratios of these methods vary according to the structure of the data used, its linearity or complexity, and the time range it covers. Forecasters generally preferred the time-series forecasting methods under the conditions where past data is available.The time-series analyses foresee that demand structure of the past also the sales amounts will continue in the future as is, based on the time-oriented changes. Most widespread ones of them are average-based techniques, exponential smoothing method, Holt's and Winter's exponential smoothing, trend-based techniques, seasonal change techniques, Box-Jenkins methods, ARMA and ARIMA.Apart from them, there are also multi-agent analyses taking place by way of analyzing numerous factors through statistical and mathematical models, in which they are not handled only within the time dimension but considering the relation between the dependent and independent variables. Simple-multiple regression, correlation analyses, and econometric models are the most used methods in the analyzing of the relations between the variables.Recently, intuitive methods like the artificial neural network, simulation, genetic algorithm, fuzzy logic, gray theory and support vector machines which have been used expressly in the complex problems, are used to forecast future.The objective of this study is to obtain accurate forecasting of the future demand for the products in the glass sector that provides substantial contributions to the economy of the country for the forthcoming period. The glass sector is a basic industry that provides many sectors like construction, automotive, home appliances, energy, food, drinks, beverages, pharmaceuticals, cosmetics, furniture, and electric-electronic sectors, in addition to its own activity field, with fundamental inputs and that creates added value for the economy of the country. Glass consists of the combination of the inputs like sand, soda, dolomite, and quartz. In the sector, all kinds of glasses obtained from the blend or cullet (broken glass pieces) by heating/melting (smelting) method and all the products obtained by subjecting them to various processes are available.In the study firstly, the concept of forecasting, reason of needing forecasting, selection of the forecasting method, and the factors that affect the demand were explained. In the continuation of the study, the concept of demand forecasting, demand forecasting application stages, principles, and methods were mentioned. The forecasting methods handled in the literature were detailed and the methods used in the previous studies were specified. Later on, the sector which the practice will be conducted was introduced and its relation with the other sectors was explained. Finally, the SWOT (Strengths, Weaknesses, Opportunities, Threats) analysis in relation with the sector was shared. With the introduction of the company in which the practice will be carried out, the analysis section was initiated. For the analysis, the product group with which forecasting will be applied was selected. The next stage after the selection of the product group, is the level where such products will be forecasted.Decision on the demand level of the operation is important in terms of deviation of the accuracy of forecasting. It is necessary to determine an intermediate level because of the reduction of the accuracy by the forecasting attempts at very low or very high levels in the product breakdown. Such levels can be in the form of stock keeping unit (SKU) or demand forecasting unit (DFU) which is an upper level. In the study, DFU level will be used. The reasons for applying forecasting at this level, are the adaptation of the sales teams to the demand designated by the customer and its follow-up, determination of products in regional separations as domestic or foreign markets, selection of the right products with regular demand, and presentation of different points that distinguish the demand. This level which is the demand forecasting unit is decided at the meeting held by the team in the sales department.The next stage after determining the level of DFU for the product groups is the determination of the region with the highest sales intensity, belonging to the glass company that is the greatest producer on the domestic market. In order to separate products at the DFU level which forecasting methods will be applied, ABC-XYZ analysis was used. Accurate classification of products is highly important in terms of stock management and planning of potential products and materials. ABC analysis conducts calculation on the periodical turnover for using the materials classification. Decision-making is ensured through the unit cost of the product and upon the consumption ratio in particular periods. With Pareto analysis principle in the A, B, and C classification parameters, the A character defines the few but highly valuable materials and the C character defines the multiple but cheap materials. The XYZ analysis that support the ABC analysis, the products and materials are selected according to the fluctuation in the consumption. X parameter represent low fluctuation or fixed-rate consumption, Y parameter represents strong or moderate fluctuation due to trends or seasonal reasons, and Z parameter represents high fluctuation that is completely irregular.Among the products selected as a result of the ABC-XYZ analysis, the product mostly demanded by the customers (sold) and another product with added value containing a secondary process were taken into account for setting an example in the study. Prior to the application of the methods, the dataset was cleaned the effective factors such as campaign, price increase and decrease, and strike period.In practice, as a method, among the time-series analyses, moving average, exponential smoothing, Holt's exponential smoothing, Winter's exponential smoothing, and ARIMA methods, among the artificial intelligence and intuitive algorithms, artificial neural networks, were used. Artificial neural networks have been developed with the inspiration of the working principles of human brain and it is a mathematical model that is flexible, able to adapt to new circumstances, applicable to complex problems, and with the competence of exchanging information and learning by experience. While linear methods are generally used in forecasting, these methods have remained insufficient some issues. YSA that yields good results especially in the solution of nonlinear complex problems is applied in many fields like forecasting, classification, learning, educating, association, and modeling, in an efficient manner.In practice, the Minitab 17 statistics software program was used for the time series methods and Neurosolutions 5.05 Developer was used together with Excel interface for the artificial neural networks. By analyzing the 36-month datasets of the selected 2 products, sensitivity analysis was conducted for each method through the model designs determined within itself. As a result, most suitable coefficient and parameters were chosen according to the error ratios. With each method determined in the light of the parameters with high accuracy rate, sales data for the 12-month forthcoming period of two products was analyzed.In the literature, many forecasting error measurements used to measure the error performance, such as the standard deviation, mean absolute deviation, mean percentage error, mean absolute percentage error, mean error square, root of the mean error square, and monitoring signal are available. What appeared as a result of the researches is that the most widespread forecasting error measurement is the mean absolute percentage error (MAPE). In the final stage of the study, forecasting performances in the soonest period of all the methods discussed were evaluated according to the mean absolute percentage error measurement and the method that yields the optimum result for forecasting future came to the light.The thing that is to be done as a result of this application is to receive the opinions of the experts. Enrichment meetings are held with the experienced persons from the sales teams on the results. The objective of such meetings is to evaluate the results of forecasting and to reflect a different result that they predict through their experiences, if any. In addition to them, some factors like economic indicators, campaign activities, and price changes are taken into account to finalize the demand forecasting. Thanks to this, decision is taken in the final forecasting results.