Ai-Powered Sales Demand Forecasting and Desicion Support system

In the last few years, retail and supply chain companies have started relying more and more on data to guide their daily decisions. Every transaction, every product movement, and every sales record is stored somewhere. Turning raw data into meaningful decisions is not always simple, even with all this information available,. One area where this difficulty becomes very noticeable is sales demand forecasting. When demand is not predicted correctly, the effects can be serious. Products may go out of stock when customers are ready to buy, or companies may end up storing large quantities of unsold goods. Both situations lead to financial loss and inefficient use of resources. Because of these challenges, improving demand forecasting methods has become an important topic of study. Many traditional forecasting methods focus mainly on past sales numbers. While historical data does provide useful insights, it does not always explain why demand changes suddenly. In reality, sales are influenced by seasonal trends, promotional activities, holidays, and unexpected market behavior. When these factors are not included in the forecasting process, the predictions may not reflect real world conditions. Another concern is that most forecasting systems present their results as numbers or graphs that require technical knowledge to interpret. Managers who are responsible for decision making may find it difficult to To deal with these limitations, this research proposes a framework that combines time series forecasting with a Decision Support System. The system collects historical sales data at the Stock Keeping Unit level from enterprise databases. Along with this data, external variables that influence demand patterns are also considered. The data is stored in PostgreSQL and then prepared carefully through preprocessing steps such as removing unusual entries, adjusting data scales, and creating meaningful features that improve prediction performance. For forecasting, the Neural Prophet model is used because it can capture overall trends, seasonal patterns, and sudden variations caused by events. However, generating predictions alone does not fully solve the problem if the results are difficult to understand. A chatbot powered by a locally deployed Large Language Model is integrated into the framework to make the system more useful in practical situations. Managers can ask questions in simple language, explore possible scenarios, and receive clear explanations instead of only numerical outputs through this interface. This approach makes the system more interactive and easier to use. Therefore, this research does not focus only on improving prediction accuracy. It also aims to make forecasting results more understandable and easily accessible for decision makers. The proposed system attempts to create a practical and scalable solution that better supports planning and strategy in retail and supply chain environments by combining analytical models with conversational interaction