Dynamic Aggregation for Time Series Forecasting

The proliferation of data has led to demand sensing and shaping at the most granular product and geography levels. This has led to a need to optimize tens and many times hundreds of millions of geography product treatments on a weekly basis. The amount of data has overwhelmed the ability to monitor individual recommendations, even by exception. In this scenario, it is imperative that the underlying demand modeling process be as stable as it is highly accurate. The methodology is geared towards automated forecasting systems with large amounts of time series inputs of varying volume and volatility. These systems are often encountered in Retail and Consumer Packaged Goods (CPG) applications such as replenishment and pricing. This paper outlines a dynamic modeling approach that produces stable and highly accurate demand forecasts.