A0626
Title: Sequential forecasting for bursty count data
Authors: Kaoru Irie - University of Tokyo (Japan) [presenting]
Aktekin Tevfik - University of New Hampshire (United States)
Chris Glynn - Zillow Group (United States)
Abstract: Existing methods for sequentially analyzing count data typically utilize a discounting strategy, where the contribution of past observations to parameter updates diminishes with elapsed time. Discount factor techniques offer an intuitive approach to sequentially updating parameters with weighted contributions from all previously observed data; however, when the time series undergoes a sudden change and the observed count significantly jumps, parameter estimates are slow to adapt, as they are heavily informed by data observed prior to the structural break. We introduce an augmented Poisson-gamma state-space (PGSS) model whose state evolution structure is flexible and responsive to sudden changes in the level of counts, focusing on consumer demand settings where sequential and online learning/forecasting are of great interest. Such adaptability is achieved by augmenting the state vector of the PGSS model with an additional state variable for a time-varying discount factor. We develop an efficient particle-based estimation procedure that is suitable for sequential analysis, allowing us to estimate dynamic state variables and static parameters via closed-form conditional sufficient statistics. To illustrate how the state-augmented PGSS model performs with data that exhibit bursts, we present results from a case study to monitor and forecast web traffic data.