Add link to IIF post

Author: ryantibs

content/blog/2021-09-30-ensemble-analysis.Rmd

@@ -27,6 +27,8 @@ output:
 knitr::opts_chunk$set(warning=FALSE, message=FALSE)
 ```
 
+*(This is cross-posted from the IIF blog; [see here](https://forecasters.org/blog/2021/09/28/on-the-predictability-of-covid-19/) for the original blog post.)*
+
 The [US COVID-19 Forecast Hub](https://covid19forecasthub.org/) has served as a central repository for forecasts of COVID-19 in since April 2020, collecting predictions of state-level cases, hospitalizations and deaths for 4 weeks into the future. We conducted an analysis that evaluated probabilistic accuracy of the ensemble forecast generated by the Hub each week. Short-term forecasts of COVID-19 deaths showed reliable performance across all horizons. However, forecasts of cases and hospitalizations showed repeated, sustained lapses in accuracy for longer-term forecasts, especially at key points during some the larger pandemic waves. Therefore, starting in September 2021, the Hub decided to suspend the inclusion of 2 through 4 week ahead case forecasts and 15 through 28 day ahead hospitalization forecasts in the official ensemble that is generated every week. Modelers and forecasters should continue to innovate and investigate so we can continue to build our understanding of how models can be used to anticipate changes in COVID trends and serve the needs of decision-makers and the general public.
 
 ## A Brief History of COVID-19 Forecasting in the US
@@ -307,4 +309,4 @@ We must not let some of the failures of forecasting impugn the many very importa
 
 Looking forward, some of the most important research that could be conducted to help improve forecasting would illustrate what data streams are able, in retrospect, to explain some of the variation and trends in COVID-19 transmission. This is a complicated and intricate problem, one that---like the Hub---will require multiple efforts from many research teams using different approaches, to hopefully converge on an answer. Some of our own [analyses in this realm](https://www.medrxiv.org/content/10.1101/2021.06.22.21259346v1) revealed that indicators derived from various digital surveillance streams can indeed offer modest improvements in case forecasting accuracy; but somewhat paradoxically, these improvements are most pronounced during periods of stability in case activity, and often break down during times of rapid increases in case trends (which is exactly when we would be looking for the most help). Much more work in this realm is needed, and eventually, we believe that these kinds of studies will be vital in improving our epidemic models in the future.
 
-Short-term COVID-19 forecasts of cases, hospitalizations, and deaths have had moments of good accuracy and also, especially for longer-term case and hospitalization forecasts, their fair share of misses. However, we know that the broader modeling community is not ready to throw in the towel. There is a lot of work to do over the coming months and years to improve these models, the forecasts, and our understanding of how to use them. So you might say that our small but blossoming field has just entered a new era, and in this sense, we are just getting started.
+Short-term COVID-19 forecasts of cases, hospitalizations, and deaths have had moments of good accuracy and also, especially for longer-term case and hospitalization forecasts, their fair share of misses. However, we know that the broader modeling community is not ready to throw in the towel. There is a lot of work to do over the coming months and years to improve these models, the forecasts, and our understanding of how to use them. So you might say that our small but blossoming field has just entered a new era, and in this sense, we are just getting started.