Прогнозування рівнів майбутніх продажів для систем планування ресурсів підприємств

Abstract

У цій статті розглянуто методи прогнозування рівнів майбутніх продажів і можливості їх використання у сучасних системах планування ресурсів підприємства. На прикладі Dynamics 365 BusinessCentral розглянуто практичне застосування таких методів, у тому числі з допомогою методів машинного навчання. Також під час роботи було досліджено наявне рішення, що базується на аналізі часових рядів (timeseries), і запропоновано доповнення із застосуванням кластерного аналізу (clustering).

This paper reviews sales forecasting methods and its ability to be used in modern enterprise resource planning systems. Having Dynamics 365 Business Central as an example was reviewed for the practical application of such methods, including those using machine learning techniques. During the described work existing solution based in time series algorithms was investigated and analyzed in detail, suggestions were made regarding the clustering techniques usage. Enterprise resource planning is defined as an ability to deliver the business software that unifies processes and data models within various areas such as finance, human resource management, supply chain management, manufacturing, distribution, etc. This paper highlights the problems that can be solved using this type of software, namely: automate the processes that were not formalized before; unify data captured by different departments for its better analysis; collect, group, and aggregate information for analytical and reporting purposes. Pros and cons for building its own solutions compared to implementation of solutions from reliable vendors are presented. Top vendors on the market are identified. Benefits of accurate sales forecasting and its importance to the business are identified: preventing missed sales, decreasing storage costs, improving marketing activities, decreasing write-offs, and improving cash flow. From another side, the common sales forecasting techniques are listed and described: clustering, descriptive, factor, time series, and regression analysis methods. Machine learning is considered as a prospective option for efficient automation of the analytical processes. Next, the specific implementation of enterprise resource planning system – Dynamics 365 Business Central – is reviewed in scope of the sales forecasting problem. It is identified that the ability of the standard solution is limited with Sales Budgeting functionality that can support manual sales forecasting but lacks automation for the future sales figures. However, its system can be extended using Sales and Inventory Forecast which in its turn can utilize the benefits of artificial intelligence techniques by means of Microsoft Azure Machine Learning service. Existing experiment that was reviewed in scope of this paper, had been based on Microsoft Time Series algorithm and combined the best results from the following techniques: autoregressive integrated moving average (ARIMA), exponential smoothing (ETS), and seasonal and trend decomposition using LOESS (STL). The following potential areas for improvement were identified in the existing approach: for large item portfolio the seasonality may significantly differ from one group of items to other; the same approach is applied to items on different life cycle stage; there could be a huge difference in behavior of slow and fast moving items; non-moving items may impact the overall picture. As a potential improvement, the usage of the clustering analysis methods is proposed. The author suggests splitting items into clusters using k-means machine learning method before performing time series analysis and assessing the impact of this approach on a large verified data set.