Re: [Manycouches] covid forecast of a (no)wave for Summer IETF in Philadelphia, USA

[Alexandre Petrescu <[e-mail]>]

For the state of Pennsylvania:

- the forecast for the next 4 weeks is a growth in number of
   cases.

   Based on their recent success rates, the selected models are: BPagano-
   RtDriven (forecasts an exponential growth), KITmetricslab-
   select_ensemble (a steady growth), MUNI-ARIMA (a steady growth)

- my empirical forecast still tells a peak on May 6th.  Probably I will
   be proved wrong, with a peak to appear a few weeks later, I dont know.
   I will know whether I was right or wrong in around 2 weeks.

For the Philadelphia County (this County is same thing as the City of 
Philadelphia):

- the forecast for the next week is a very slight growth in the number
   of cases.

   Only the model COVIDhub-ensemble is available for Philadelphia; the
   other models available individually for Pennsylvania are not available
   individually for Philadelphia, even if they are 'averaged' in the
   'ensemble' model.

   This model COVIDhub-ensemble has proven to be optimistic in the recent
   past, i.e. it predicted less cases than the truth.

   Ideally, one would take the source code of the selected models for
   Pennsylvania data  (see above the model names) and apply them
   individually for Philadelphia data, to obtain a better forecast for
   this Philadelphia.

- my empirical forecast still tells a peak on May 15th.

   This is obtained if fitting a line through the surprisingly constant
   set of 3 equal values 133, 133, 133! (133 is the number of days
   between the observed peaks in Philadelphia : Nov. 29th, 2020, April
   11th, 2021, Aug. 22nd, 2021, Jan. 2nd, 2022, see the orange vertical
   bars at 
https://www.phila.gov/programs/coronavirus-disease-2019-covid-19/data/testing/ 
), and ignoring the first peak of April 5th, 2020,
   situated at 238 days from the second peak.  The reason of potential
   ignoring of this first peak is that measurements changed significantly
   in the first part of the pandemy, everywhere.

   If this happens (a 4th time a difference between peaks of 133 days)
   then I will have strong reason to doubt the data on phila.gov site.

   If on the other had one does not ignore the date of the first peak,
   then a parabolic could be fit (rather than a linear); this would
   forecast the next peak would be on August 1st, 2022 (at 211 days from
   Jan . 2nd, 2022, the last peak).

   As one can see, the range between those two forecasts (May 15th and
   August 1st) is way too wide to be of any use.

   I still look for an empirical method to predict the next peak for
   Philadelphia.

Alex

Le 26/04/2022 à 15:41, Alexandre Petrescu a écrit :
> For the state of Pennsylvania:
> - again today, the wave of new cases in Pennsylvania is growing a little
>    bit higher than forecasted a week ago by the model COVIDhub-ensemble
>    (group of models).  The forecasts of the individual models Microsoft-
>    DeepSTIA, SDSC_ISG-TrendModel and MUNI-ARIMA are closest to the
>    measured number of cases (the 'truth').
> - by MUNI-ARIMA and Microsoft-DeepSTIA models, there is going to be
>    growth in the number of cases until or beyond May 15th.
> - my empirical calculations, which are not models, is reminded, said
>    that May 6th will be a peak.
> 
> For the Philadelphia County:
> - the model COVIDhub-ensemble predicts from Apr. 16th a growth of
>    approximately 80 new cases on April 23rd.  We'll learn in a few days
>    how correct that prediction was, when measurements are available.
> 
> For the City of Philadelphia (is it the same as ~ County, covid-wise?):
> - I still couldnt find forecast models, be them epidemilogical or
>    emprical, that take into account the history data of the City of
>    Philadelphia.
> - my empirical calculations, is reminded, said the following:  the next
>    wave peak could be on May 15th if fitting a linear through only the
>    last 3 peak date differences, or on July 26th if fitting a parabolic
>    curve on all 4 differences.  Because, surprisingly, the past waves
>    were distanced like this: 231, 133, 133, 133, as exhibited by the
>    curve of  'Positives' at this URL:
> https://www.phila.gov/programs/coronavirus-disease-2019-covid-19/data/testing/ 
> 
> 
> Note: I was asked whether covid models take weather into account.  The 
> questioner feels natural that if weather is humid, or cold, then there 
> is more covid transmission.  I think that there are many weather 
> forecasts that are very good but I could not find a covid model that 
> takes weather forecasts into account, neither vice-versa.  There is 
> however one covid model ('SDSC_ISG-TrendModel')  that takes 'LOESS 
> seasonal decomposition' into account, where'season' == periodicity to be 
> identified, not necessarily weather, but it could be related.  This 
> SDSC_ISG_TrendModel gave a good forecast for Pennsylvania State at the 
> last run.
> 
> Note2: I was told that forecasts of covid models predict everything and 
> their contrary, and as such they should not be trusted.  My answer is 
> that indeed different models differ largely in their forecasts, but they 
> also differ in their past behaviour.  It is based on that past behaviour 
> that one should use the best model.  Ignoring the forecasts is more 
> risky; we planned for group lunch 2 times since Christmas, and cancelled 
> each time at the last minute because of covid.  Were we to use a covid 
> forecast we'd cancel less.  There are many websites that help in 
> evaluating the models, for example this one:
> https://delphi.cmu.edu/forecast-eval/
> 
> Alex
> 
> Le 20/04/2022 à 17:46, Alexandre Petrescu a écrit :
>> Hi, SHMOO,
>>
>> I am trying to produce a forecast for the covid wave in July in 
>> Philadelphia.  I am not an epidemiologist, so my forecast is only 
>> empirical.
>>
>> These are the weekly notes:
>>
>> For Pennsylvania state:
>>
>> - it might be that a small covid wave is approaching to peak in the
>>    coming 4 weeks.
>>    The measured number of new covid cases is a bit higher, compared to
>>    what the 1-week forecast said one week ago.  The measure 8131 was
>>    made on April 16th, 2022; the forecast 5599 was made on April 9th,
>>    by the 1-week forecast model 'COVIDhub-ensemble' for Pennsylvania
>>    state.
>>    (the model 'MUNI-ARIMA' forecasted 8002 which is very close this
>>    time).
>> - the model 'COVIDhub-ensemble' forecasts 8130 for April 23rd.  Only on
>>    April 27th will we learn how right that forecast was.
>> - the model 'COVIDhub-4_week_ensemble' forecasts 5656 for May 7th.
>> - my empirical calculations (root function curve through last 4 peak
>>    date differences) tells May 6th to be a next peak.
>>
>> For City of Philadelphia:
>>
>> - my empirical calculations: the next wave peak
>>    could be on May 15th if fitting a linear through only the last 3 peak
>>    date differences, or on July 26th if fitting a parabolic curve on all
>>    4 differences.  Because, surprisingly, the past waves were distanced
>>    like this: 231, 133, 133, 133, as exhibited by the curve of
>>    'Positives' at this URL
>>
>> https://www.phila.gov/programs/coronavirus-disease-2019-covid-19/data/testing/ 
>>
>>
>> - I have not found other forecasts specific to the City of Philadelphia,
>>    or to the Philadelphia County.
>>
>> For the planner, there are some dashboards to play with hypothesis of 
>> lockdowns, masks promotion, vaccination, variants, waning or escape:
>> https://analytics-tools.shinyapps.io/covid19simulator04/
>> https://covid19scenariomodelinghub.org/viz.html
>>
>> Alex
>>
>> Le 15/04/2022 à 12:46, Alexandre Petrescu a écrit :
>>> Hi, Jason,
>>>
>>> Le 14/04/2022 à 16:30, Livingood, Jason a écrit :
>>>> On 4/14/22, 07:22, "Manycouches on behalf of Alexandre 
>>>> Petrescu"<[e-mail] on behalf of 
>>>> [e-mail]>  wrote:
>>>>> https://covid19.healthdata.org/united-states-of-america/pennsylvania?view=cumulative-deaths&tab=trend 
>>>>>
>>>> I would not look at a Pennsylvania dashboard, as the state of quite 
>>>> large and extremely varied in local support for COVID mitigation 
>>>> measures & vaccination rates. As you might imagine, large metro 
>>>> areas tend to diverge from the rural parts of the state.
>>>>
>>>> You can find Philadelphia's stats 
>>>> athttps://www.phila.gov/programs/coronavirus-disease-2019-covid-19/data/testing/. 
>>>> The city is reimposing an indoor mask mandate in the next few days 
>>>> (though positive tests are still quite low vs. the January 2022 
>>>> Omicron BA.1 surge.
>>>
>>> Thank you for the URL.  More local data is indeed more valuable.
>>>
>>> I might use it for my own calculations, which would be more precise, 
>>> would offer more confidence.
>>>
>>>> In any case, what C19 will look like in Philadelphia or anywhere 
>>>> else in July 2022 is anyone's guess.
>>>
>>> I agree it is anyone's guess; but I would not discard it, because it 
>>> might be the only way to try to see better.
>>>
>>> Other than my own guess yet to be formed, there are these forecasts 
>>> on the Internet; their forecast for July 23rd depends on a 'model', a 
>>> 'certainty', and a 'scenario'.
>>>
>>> If I were to select only the model Ensemble, certainty 95% and 
>>> scenario 'pessimistic but no immune variant', then their forecast of 
>>> March 19th says that on July 23rd the state of Pennsylvania would be 
>>> right after a peak of new cases.
>>>
>>> This is the illustration from this URL 
>>> https://covid19scenariomodelinghub.org/viz.html  (the vertical date 
>>> bar is drawned by me):
>>>
>>>
>>>
>>> Alex
>>>
>>>
>>>> Jason
>>>>
>>>>
>>>
>>> _______________________________________________
>>> Manycouches mailing list
>>> [e-mail]
>>> https://www.ietf.org/mailman/listinfo/manycouches
>>
>> _______________________________________________
>> Manycouches mailing list
>> [e-mail]
>> https://www.ietf.org/mailman/listinfo/manycouches
> 
> _______________________________________________
> Manycouches mailing list
> [e-mail]
> https://www.ietf.org/mailman/listinfo/manycouches