Forecasting new product diffusion with agent-based models

Agent-based model (ABM) has been widely used to explore the influence of complex interactions and individual heterogeneity on the diffusion of innovation, while it is seldom used as a forecasting tool in the innovation diffusion literature. This paper introduces a novel approach of forecasting new product diffusion with ABMs. The ABM is built on the hidden influence network (HIN) over which the innovation diffuses. An efficient method is presented to estimate non-structural parameters (i.e., p, q and m) and a multinomial logistic model is formulated to identify the type of the HIN for diffusion data. The simulation study shows that the trained logistic model performs well in inferring the HINs for most simulated diffusion data sets but poorly for those generated by ABMs with similar HINs. Therefore, to reduce the possible prediction loss arising from the misspecification of the HIN, three methods, namely, the predicted HIN, the weighted averaging and simple averaging, are developed to forecast new products diffusion. Their performances are evaluated by using a data set composed of 317 time series on consumer durables penetration. The results show that most identified HINs have moderate topology, and that our methods outperform four classical differential equation based diffusion models in both short-term and long-term prediction.