COVID-19 Herd Immunity Forecasting with N-BEATS

A historical 2021 N-BEATS study of Indonesia's COVID-19 cases and vaccinations. Reported R-squared values are in-sample checks, not held-out tests.

0.992
Daily case model R-squared, in-sample historical forecast
0.9995
Full-dose vaccination model R-squared, in-sample historical forecast
208 days
Daily data used, from 27 January 2021

Role: Co-author and notebook author in a three-person team
Period: Aug 2021
Project type: Featured
Room in the 3D HQ: L3:covid-herd-immunity-forecast

Results

0.992
Daily case model R-squared, in-sample historical forecast
Historical forecast from 20 March to 22 August 2021 on the training series: an in-sample check, not a held-out test (Darts r2_score, notebook output).
0.9995
Full-dose vaccination model R-squared, in-sample historical forecast
Same in-sample historical forecast method on cumulative full-dose vaccinations (notebook output).
208 days
Daily data used, from 27 January 2021
Notebook output of the main case model.
About 10 months
To 80% full vaccination at the August 2021 pace
At an average of 273,318 second doses a day (the paper in the repository).

Problem

In August 2021 Indonesia was coming off its largest COVID-19 wave, and the question was when cases would ease and how long full vaccination toward herd immunity would take at the current pace.

Contribution

I combined daily case, recovery and vaccination data from January 2021 with population figures, mobility data and weights for policies and holidays.
I trained N-BEATS models in Darts for daily cases (with mobility as a covariate) and for full-dose vaccinations, and backtested them with historical forecasts from March 2021.
The team turned the forecasts into estimates of when cases would bottom out and how long 80% full vaccination would take, published at IEEE ICICyTA 2021.