Ensembles of Nearest Neighbor Forecasts

Nearest neighbor forecasting models are attractive with their simplicity and the ability to predict complex nonlinear behavior. They rely on the assumption that observations similar to the target one are also likely to have similar outcomes. A common practice in nearest neighbor model selection is to compute the globally optimal number of neighbors on a validation set, which is later applied for all incoming queries. For certain queries, however, this number may be suboptimal and forecasts that deviate a lot from the true realization could be produced. To address the problem we propose an alternative approach of training ensembles of nearest neighbor predictors that determine the best number of neighbors for individual queries. We demonstrate that the forecasts of the ensembles improve significantly on the globally optimal single predictors.