Dealer based demand forecasting application in automotive industry

In today's globalized world, it has become harder to survive for a long time in global market for every company; especially for the production companies. As they use limited sources to manufacture the products, the whole process needs to be managed carefully from strategic planning to after sales services at production companies. Wasting the limited sources by having unefficient process, producing non-demanding products, etc. will cause trouble and company may lose its competitiveness. To prevent the wasting from the perspective of producing the right products will help companies to manage the stocks and reducing the consumption of the limited sources inefficiently. These kind of problems drive the companies to make the studies and be able to predict the future more accurate. There are lots of methods to predict the future, companies may choose one or more of them inline with their targets and company structures to predict their future and take the necessary actions accordingly. The methods have become more accurate and data driven with the technological developments. Companies may also focus on different areas to improve inline with their strategies like customer satisfaction, cost reduction, stock management etc. It also affects the chosing of forecasting method. Production companies need to take into the consideration of demand and should plan the production accordingly. The data based forecasting methods are being used generally for the demand forecasting. Company A which sells passenger and commercial vehicles to its customers via its dealers has been investigated at this thesis. Company A is a company that wants to improve the demand forecasting accuracy on dealer base. Model B which is a member of the passenger car portfolio has been choosen as the pilot product.There are 85 dealers in Turkey and 30 vehicle combinations that can be ordered. At current status, the dealers order the cars by using their experience. This method works but needs to be improved as the dealers have shortage for best seller vehicle which may cause losing the customer. On the other hand, there are lots of vehicles at the stocks for the long time which customers do not prefer to buy. In addition to these status, the passenger cars are the export cars and have long delivery durations once they are ordered. Company A wants a forecasting model that can guide the dealers for their future orders. So Company A has provided the sales data of Model B for and the literature research has been conducted to set up the model . Firstly, as the starting point of the study the prediction methods have been investigated in detail. There are 2 main categories in prediction methods that called as qualitative and quantitive. In qualitative methods, personal experience and expertise are the most important things. These methods' results change from person to person inline with their knowledge and experiences. It is important to choose theright person or people for the forecasting by using qualitative methods otherwise it might direct wrongly. In quantitive methods, the predictions based on the data sets and mathematical equations to eliminate the affect of personality and subjectivity on the forecasting. These methods also require personel knowledge but it is required to use the methods. Personal thoughts do not affect the model. Artificial intelligence method has been investigated as a part of quantitive methods in this thesis. Secondly, the dynamics and the trends in automotive industry has been added into the study to understand the drivers of customers buying decisions. There are macro and micro elements that need to be considered to have an accurate forecasting result. In Turkey, the drivers of the buying factors are diffent from most of the countries in Europe as Turkey is one of the developing countries. he factors that affect customers buying decisions in Turkey are the price of the cars, the income level, gasoline cost, tax policy for cars, interest rate at credits, the uncertainty of the future because of the non-stabil exchange rates. Having the non-stabil exchange rates from year to year makes the customers uncomfortable. When the exchange rate is changed, which is mostly in bad way, the customers become poor from year to year. So they defer theirbuying decisions or they change their minds not to buy a vehicle. The factors and possible resolutions are being mentioned in Chapter 3. The average age of the car park in Europe and Turkey is also different, the average age of the registered vehicles is between 7 and 10 in Europe however it has been calculated as 12.4 in Turkey. 21% of the total vehicles are older than 20 years. These numbers may be interpreted as Turkey is a growing market from car sales perspective also. Artifical neural networks is being investigated in Chapter 4. The definition of the artifical neural network, the components of the network, the training methods, the structure of the networks etc are being mentioned in detail. The aim of this section is to give a common highlight about the working principles of neural networks. The advantages and disadvantages comparing with other prediction tools can be also found in this section. Artificial networks are one of the quantitive methods that allow the user to set up a structure for forecasting based on the historical data by determining learning function, momentum coefficient, learning coefficient etc. In the application, the data set that was provided by Company A sales planning team has been used. Data set contains Model B vehicle sales from January 2014 to June 2017. There are the information of the dealers, the regions, color of the car, campaign data in monthy basis etc in the file. All of the information about the each sale could not be shared due to the personel data prevention law. The sales, the color of the car and the region info have been used to set the structure of the study. The seasonality affect was not considered as the campaign data has been added as an category input to the model. To predict the sales at time t, the t-1, t-2,..., t-12 data sets are being used as input. The alternatives have been investigated to set up a structure and the decision was to use Matlab program. The model has been run at Matlab R2017b with Neural Networks tool. The learning ratio, momentum ratio, epoch etc were found with several trials. The optimum combination has been choosen as the model inputs. All details can be found at Chapter 5. The forecasting methods are not 100% accurate, there is always a margin of error. The methods' accuracy can be calculated by comparing the model output and the real data. The result of this comparison gives an idea about the reliability of model. Thereare different ways to figure out the margin of error like mean squared error, root mean squared error, mean absolute error etc. MSE (Mean Squared Error) method has been used to evaluate the confidence of the study. The MSE of the model has been calculated as %0.67 which means the results have high accuracy. The model input belongs to top 1 selling vehicle of the Company A. The other models of this category will be studied with the Company A afterwards. The model can be extended to include the commercial vehicle category after the passenger car study completion. After calculating the margin of error and having the high accurate outputs, the recommendation has been shared with Company A to use an artifical neural network for the forecasting the demand of Model B. Working with one of the experts in the artifical intelligence area will help to company to have more accurate demand forecasting and manage the stock level in parallel. Stock control is not the scope of this study but it will be improved after the implementation of the network as a natural output. Currently the undesired vehicles have stock aging at dealers' stocks with no profit but extra cost. Long stock aging causes the the model year change of the vehicles and unable to sell them without discount. With the implementation of this model, company will forecast the demand with a scientific method and have the optimum mix of the vehicles in their stocks. The another recommendation has been shared with Company A to have a set up a common car pool that allows the company to continue offering the wide range of products to its customers. This study showed that there are 8-10 vehicle combinations genearate 80% of the sales. The other vehicle combinations which are the less demanding ones could be collected in a pool that Company A owns and all the dealers will be able to order from this pool if the customer wants that variant. The company has kicked off the process and the reviews with the suppliers are ongoing inline with the thesis output.