Time Series Forecasting And Big Data: Can An Ensemble Of Weak Learners Decrease Costs While Maintaining Accuracy?

While the main focus of research for time series forecasting models has been on improving the accuracy of the forecasts by increasingly complex models and parameter tuning, the increasingly large data sets being used require constantly stronger hardware and computational resources. Typically, the accuracy of models during training is evaluated by looking at in-sample errors, which does not always translate into better out-of-sample forecasts. In this paper, we will investigate whether possibly combining weaker models can result in acceptable results especially when compared to more resource-intensive parameter optimizations. We then take a look at the trade-off between this race for finding the best model and balancing the resource cost from a monetary perspective.