Evaluation Procedures for Forecasting with Spatiotemporal Data

The increasing use of sensor networks has led to an ever larger number available spatiotemporal datasets. Forecasting applications using this type data are frequently motivated by important domains such as environmental monitoring. Being able properly assess the performance different forecasting approaches is fundamental achieve progress. However, traditional estimation procedures, cross-validation, face challenges due implicit dependence between observations in In paper, we empirically compare several variants cross-validation (CV) and out-of-sample (OOS) both artificially generated real-world Our results show CV OOS reporting useful estimates, but they suggest that blocking space and/or time may be mitigating CV’s bias underestimate error. Overall, our study shows importance considering dependencies when estimating models.