Just released my COVID English Hospitalisations Explorer & Forecaster http://shiny2.ma.ic.ac.uk/users/gnason/  Online app using data direct from @NHSEngland Enables forecasting FROM a specific date a number of days ahead that you choose. @d_spiegel #covid #covid19 @BristOliver @rowlsmanthorpe

1. Your choice of prediction intervals and you can easily download forecasts and prediction intervals in a format of your choice.

2. Additional exploration of English hospitalizations series in terms of trend, non-stationarity detection and spectrum to identify cycles.

3. FULLY documented, with guidance notes. Intended to be fully reproducible identifying data source, tried and trusted methods, package dependence and full source code. Please let me know if there's anything missing.

4. [I still pretty sure I cannot reproduce Government forecasts/projections as they are not documented FULLY. This is basic and should be insisted on @StatsRegulation ]

5. Here's a little example: go to the app http://shiny2.ma.ic.ac.uk/users/gnason/  and follow these instructions. Hit "Use Latest NHS Data" (on mine the Last Date changes to 10th November)

6. Now change forecast horizon slider from 7 to 14 days. Change the Display Start Date to, e.g., 1st Oct. All data still being used, just displaying more recent data. You will see the point and interval forecasts on the far right.

7. On my plot (and yours too) hospitalization forecasts seem relatively flat. That's good news! Let's compare to what the Government forecasted on 26th Oct.

8. Change the "Model End Date" to 26th Oct. We're not pretending that today is 26th Oct. The forecaster predicts steady increase in hospitalizations (bad!). Let's see how this compares to the Government forecast. Click "+No10 projections"

9. The Gov. projections from 26th Oct are added in red with the lower- and upper-quartiles. It is interesting to see how close the Gov projections are to the app's forecasts and prediction intervals.

10. So, here you have two largely independent forecasting systems, using a different statistical basis, predicting similar things.

11. But the shiny app is using "standard" methods on a single time series, whereas the Gov projections are a (slightly mysterious) blend of the outputs from six different models produced by large, resourced, teams from nine different institutions.

12. How did we both do? Now click "Plot Actual Hosp beyond Model End Date" and the no. of hospitalizations are added. It's really great that the number of hospitalizations were less than forecast (by BOTH). Now perhaps running at 250 fewer per day?

13. But actual hospitalizations remain, more or less around the lower-quartile prediction/projection interval. So, "within tolerances".

14. Why weren't the forecasts better? Well, change the Model End Date now to 27th Oct. We're now pretending that we're doing the modelling/forecasting one day later than before.

15. I don't have the Gov forecasts made on that day (if they exist?). However, the shiny app forecasts are now much closer to what actually happened. Same with 25th, 27th, 28th.

16. So, what gets forecasted really does depend on what day you make the forecast. Were there other Gov forecasts publicised? Were decisions based on one day's forecasting? Surely not?

17. Will look at cases and deaths in due course.