Dynamic Pattern Matching Using Temporal Data Mining for Demand Forecasting

Traditional time series methods are designed to analyze historical data and develop models to explain the observed behaviors and then predict future value(s) through the extrapolation from the models. The underlying premise is that the future values should follow the path of the historical data analyzed by the time series methods, and as such, these methods necessitate a significant amount of historical data to validate the model. However, this assumption may not make sense for applications, such as demand forecasting, where the characteristics of the time series may alter frequently because of the changes of consumers’ behavior and/or cooperate strategies such as promotions. As the product life cycle gets shorter as it tends to be in today’s e-business, it becomes increasingly difficult to make a forecast using traditional time series methods. In response to this challenge, this paper proposes a novel pattern matching procedure to decide whether one or combination of several patterns actually represents the development of the time series and then to use the patterns in forecasting. Several pattern transformation algorithms are also proposed to facilitate a flexible match. Rematching through dynamic reevaluation of the new data may be needed until the true development of the time series is discovered. Initial evaluation indicates superior performance in predicting the demand of a new product