Mixed membership nearest neighbor model with feature difference

Abstract The nearest neighbor model has been a popular for spatial data. This assumes distance‐based neighborhood structure among set of entities that define the observations in dataset. Values dependent variable are function corresponding values its neighbors and own explanatory variables. In this paper, we extend concept “spatial neighbors” to more general network dependency where is not necessarily distance‐based. As each entity can belong multiple groups, be related by networks. By exploring overlapping information exhibited an neighbors, modify traditional autoregressive with disturbances (SARAR) three ways. First, networks allowed. Second, differences between variables their neighboring averages applied. Third, apart from independent innovations, investigate case innovations also multiple‐network dependent. Statistical inference parameters achieved via maximizing profile log‐likelihood function. standard errors maximum likelihood estimators calculated inverse Fisher matrix. Four variations models applied predict changes stock prices 555 companies listed Hong Kong Stock Exchange. attribute differencing produces smaller out‐of‐sample mean prediction error sum squares than other models.