Expand conclusions (#49)

analysis/paper/hubEnsembles_manuscript.qmd

@@ -1473,7 +1473,13 @@ The choice of an appropriate ensemble aggregation method may depend on the forec
 
 ## Summary and discussion {#sec-conclusions}
 
-Ensembles of independent models are a powerful tool to generate more accurate and more reliable forecasts of future outcomes than a single model alone. Here, we have demonstrated how to utilize [hubEnsembles]{.pkg}, a simple and flexible framework to combine individual model forecasts and create ensemble predictions. When using [hubEnsembles]{.pkg}, it is important to carefully choose an ensemble method that is well suited for the situation. Although there may not be a universal "best" method, matching the properties of a given ensemble method with the features of the component models will likely yield best results. For example, we showed for forecasts of seasonal influenza in the US, the quantile median ensemble performed best overall, but the linear pool method had advantages during periods of rapid change, when outlying component forecasts were likely more important. Notably, all ensemble methods outperformed the baseline model. These performance improvements from ensemble models motivate the use of a "hub-based" approach to prediction for infectious diseases and in other fields. Fitting within the larger suite of "hubverse" tools that support such efforts, the [hubEnsembles]{.pkg} package provides important software infrastructure for leveraging the power of multi-model ensembles.
+Ensembles of independent models are a powerful tool to generate more accurate and more reliable predictions of future outcomes than a single model alone. Here, we have demonstrated how to utilize [hubEnsembles]{.pkg}, a simple and flexible framework to combine individual model predictions into an ensemble.
+
+The [hubEnsembles]{.pkg} package is situated within the larger hubverse collection of open-source software and data tools to support collaborative modeling exercises. Collaborative hubs offer many benefits, including serving as a centralized entity to guide and elicit predictions from multiple independent models [@reich2022]. Given the increasing popularity of multi-model ensembles and collaborative hubs, there is a clear need for generalized data standards and software infrastructure to support these hubs. By addressing this need, the hubverse suite of tools can reduce duplicative efforts across existing hubs, support other communities engaged in collaborative efforts, and enable the adoption of multi-model approaches in new domains.
+
+When using [hubEnsembles]{.pkg}, it is important to carefully choose an ensemble method that is well suited for the situation. Although there may not be a universal "best" method, matching the properties of a given ensemble method with the features of the component models will likely yield best results [@howerton2023]. Our case study on seasonal influenza forecasts in the US demonstrates this point. The quantile median ensemble performed best overall for a range of metrics, including weighted interval score, mean absolute error, and prediction interval coverage. Yet, the linear pool method, which generates an ensemble with wider prediction intervals, demonstrated performance advantages during periods of rapid change, when outlying component forecasts were likely more important. Notably, all ensemble methods outperformed the baseline model. The performance improvements from ensemble models motivate the use of a "hub-based" approach to prediction for infectious diseases and in other fields.
+
+Ongoing development of the [hubEnsembles]{.pkg} package and the larger suite of hubverse tools will continue to support multi-model predictions in new ways, including for example supporting additional types of predictions and enabling cloud-based data storage. All such infrastructure will ultimately provide a comprehensive suite of open-source software tools for leveraging the power of collaborative hubs and multi-model ensembles. 
 
 ## Acknowledgements {.unnumbered}