A0905
Title: Online model averaging prediction
Authors: Jun Liao - Renmin University of China (China) [presenting]
Abstract: For massive data, it is often difficult to describe the characteristics of the data by a single model, and hence, the associated prediction may not be desirable due to the model misspecification. The online model averaging predictions for massive data is developed, which is particularly appropriate for the common scenario where the prediction model may be misspecified, and the large-scale data are collected sequentially. The proposed methods do not need to use the whole data of the individual level and only need to utilize the current batch of data and some statistics based on the previous batches. Also, the new methods adapt to both the continuous response and discrete response cases, which include the common regression models, such as ordinary linear regression, nonlinear regression, logistic regression, and Poisson regression. The online model averaging estimators are shown to be asymptotically optimal. Further, the convergence rate of the weight estimator developed in terms of the squared prediction error is derived. The simulation study and real data analysis reveal that the proposed methods have the desirable finite sample performance.
