A nearest neighbor model for forecasting market response

Researchers in marketing often are interested in modeling time series and causal relationships simultaneously. The prevailing approach to doing so is a transfer function model that combines a Box-Jenkins model with regression analysis. The Box-Jenkins component assumes that a stationary, stochastic process generates each data point in the time series. We introduce a multivariate methodology that uses a nearest neighbor technique to represent time series behavior that is complex and nonstationary. This methodology represents a deterministic approach to modeling a time series as a discrete dynamic system. In this paper we describe how a time series may exhibit chaotic behavior, and present a multivariate nearest neighbor method capable of representing such behavior. We provide an empirical demonstration using store scanner data for a consumer packaged good.