Agent-based demand forecasting for supply chain management

This dissertation introduces a new agent-based demand forecasting system, which incorporates causal information in addition to historical demand data, to improve the forecasting accuracy, especially for e-commerce and supply-chain oriented businesses. Due to the lack of historical data resulting from shortened product life cycles, proliferation of products, and shifting of consumer behaviors, the ability to make accurate and consistent demand forecasts using traditional forecasting methods has been challenged. The objectives of this study are (1) to develop a forecast system to overcome limitations of traditional demand forecasting methods, (2) to utilize existing forecasting methods in a more effective way, and (3) to improve forecast accuracy through the inclusion of influential information available from both internal and external data sources.; The agent-based forecasting system is implemented through the use of four types of agents: the coordination agent, the task agent, the data collection agent, and the interface agent. The coordination agent provides weighted overall prediction from the forecasts reported by various task agents. The task agents perform forecasting by using different forecasting methods or incorporating special event patterns and possible causal effects for demand variation. The data collection agents are controlled by the coordination agent for demand data gathering. Finally, the interface agent communicates between the coordination agent and the human user. Through the periodical re-evaluation of weights, the agent-based forecasting system is able to provide improved forecasts and make adjustments when the characteristics of the demand pattern change over time.; Also, a novel combination of a pattern-matching algorithm and the weighting approach is developed as an alternative to traditional time series forecasting for limited demand data situations. The objective of the pattern-matching algorithm is to consider if certain event patterns would actually represent the immediate development of the on-going time series and to use the candidate patterns as indicators of future trends. This was shown to work well in situations with small amount of historical data, such as in the early introductory stages of a new product.; The superiority of using the multiple agent based forecast system over traditional forecasting methods is demonstrated through four case studies. In all four cases, the agent-based forecasting system was able to provide more accurate forecast than traditional time series and causal forecasting methods. The results showed statistically significant improvement over existing methods.