[Discuss] Poland parameter forecast gives unexpected increased Rt

[Inglezos]

Summary

The forecast analysis (with delay=recovered period) for Poland gives unexpected increased Rt value which results in increased cases, while actually the situation in Poland seems to stabilize:

Codes

import covsirphy as cs

data_loader = cs.DataLoader(directory="kaggle/input")
jhu_data = data_loader.jhu()
population_data = data_loader.population()
oxcgrt_data = data_loader.oxcgrt()

 x_scenario = cs.Scenario(jhu_data, population_data, "Poland")
 x_scenario.records_diff()
 x_scenario.records().tail()
 _ = x_scenario.trend()
 x_scenario.estimate(cs.SIRF)
 x_scenario.add(name="Main", days=30)
 x_scenario.fit_predict(oxcgrt_data=oxcgrt_data, name="Forecast").summary(name="Forecast")
 x_sim_df = x_scenario.simulate(name="Main")
 x_sim_df = x_scenario.simulate(name="Forecast")
 x_scenario.summary()

Environment

• CovsirPhy version: 2.14.0-delta
• Python version: 3.8
• Installation: conda/pip and Google Colab (v2.14.0)
• System: Windows

[lisphilar]

The table indicates rho sharply decreased and theta + simga increased as expected with Rt definition of SIR-F model. (figures: Scenario.history("rho"). etc.)

Were OxCGRT scores sharply changed in Poland?

[Inglezos]

It is strange because the Main scenario is totally different and seems more normal:

[Inglezos]

No, OxCGRT scores are stable:

[lisphilar]

The 7th column shows "1, 1, 1,.... 2 (31Dec2020), 2, 2, 2, 2, 1, 1...." Scenario.fit() may be a clue. (I will check the details UTC-tomorrow.)

[Inglezos]

Yes I know, but this little change (from 1 -> 2) seems improbable to lead to such a rapid increase of cases in the next days. It is about doubling the cases in just two weeks. On the contrary, since a partial score increased (stricter), then the cases should not increase that much but decrease instead, by intuition right?

[Inglezos]

Also weird forecast results seem to have Russia and Switzerland as well.

[lisphilar]

this little change (from 1 -> 2) seems improbable to lead to such a rapid increase of cases in the next days. It is about doubling the cases in just two weeks

Max values of OxCGRT scores (except stringency_index) are 3 or 4 and this not a small change.

since a partial score increased (stricter), then the cases should not increase that much but decrease instead

Yes. (stricted policy measures -> decreased rho -> decreased number of infected cases)

[lisphilar]

Details of OxCGRT indicators: https://covid19datahub.io/articles/doc/data.html https://github.com/OxCGRT/covid-policy-tracker/blob/master/documentation/codebook.md

For Poland,

• predicted sigma value was negative value.
• Cancel_events increase -> rho increase (confirmed with [15] un-expected)
• Internal_movement_restrictions increase -> rho increase (confirmed with [15] un-expected)

https://gist.github.com/lisphilar/4f71b3e62448fc0d50eed983abd3fc2a

[Inglezos]

So what is happening, what the results mean? What went wrong? As I remember, the parameters must never be negative. And since the measures changed to stricter ones, then the cases should decrease, not increase. In the above notebook analysis, the forecasted Rt is negative and equal to the unexpected value -3.83 (10Jan21- 9Jan21) !?

[lisphilar]

At the current version 2.15.0-eta-fu1, the situation was improved with "delay=12" (estiamted with .estimate_delay()), but kappa is still negative value. https://gist.github.com/lisphilar/d3ee1cd2cdfe85aaa3c3c2458326897b

Restriction to (0, 1) could be implemented, but no ideas for codes at this time.

[lisphilar]

Version 2.16.2-kappa-fu2: theta < 0 https://gist.github.com/lisphilar/190d9583a10a3c6c8678ce065a0c693b

[lisphilar]

This will be discussed in #668.