k-Nearest Neighbor Regressors Optimized by using Random Search

This work proposes a method for forecasting time series based on a model selection of k NN regressors. Our technique is simple but powerful, we propose to compose a single configuration space joining both time series parameters and k NN parameters, with the idea of performing a coupled global optimization of all parameters; then, we select a competitive model over that search space using random search and a cross-validation scheme. Our experimental results show that this strategy outperforms other complex approaches like Nearest Neighbor tuned by differential evolution (NNDE) or the Fuzzy Nearest Neighbor (FNN).