This project is performed on behalf of SIG. SIG is a global company that produces packages and filling machines for food and beverages. SIG has multiple different clients within Europe and Asia. For each of these customers, SIG customizes the package in terms of size and print. Because of this customization, SIG has a lot of different “products” (combination of a customer and one of his requested package materials) to fulfil. For the sales and operation planning purpose, SIG’s salespersons need to hand in the forecast of demand for different products every month. However, the forecasts made by the salesperson are not accurate enough and therefore the planning has to be adjusted during the month, which costs money. Therefore, this project is helping SIG forecast the sales of these customized customer package materials. In this report, the focus is placed on the customers of SIG in Thailand. However, the final delivered model can also be used for customers from other countries.
The data set came from SIG, After data processing and filtering by country Thailand, SIG has 20 customers in Thailand and 85 different “products” (combination of a customer and one of his requested package materials) of time series in total. Before the actual forecasting took place, the data is analyzed and the customer packages are categorized into three categories: active customers, past customers, and new customers. New customers are customers started to have demand for their package in the last 2 years, past customers are
customers that stopped having demand for their package in the last 2 years and active customers are customer packages that already have demand for more than 2 years and had demanded in the last 2 years as well. Only for the active customers' forecasts are given for the upcoming 12 months. For the remaining two categories, either they don’t have any demand anymore so forecasts are not necessary or there is not enough data available to make accurate forecasts.
To forecast the sales of the customized customer packages for active customers, two different models are built. A quick model is built using moving average with a window of 12 periods. Next, the best model is built for every customer package pair. This best model is the best out of five different forecasting models and is chosen based on the lowest RMSE and the forecast error distribution. The five used models are; seasonal naive, moving average, Arima, exponential smoothing and linear regression. Both the quick and the best model can give a forecast for the upcoming 12 months, where the best model can give more accurate forecasts and the quick model can give forecasts in less time.
Lastly, the recommendation is given to SIG about how they can improve the accuracy of the forecasts of every customer package. We believe this accuracy can be increased by using the forecasts that are currently given by someone of the sales department for the upcoming 12 months as external data into the automated models we build.