Published in Economic & social issues · 30 May 2020
Release date: May 2020
Coronavirus (Covid-19) forecast slide decks are provided for the first wave below. The forecast ignores the lockdowns, the reasoning being that the lockdowns would simply move cases and fatalities forward and that these numbers would eventually catch up to the forecast over a long forecast range of 365 days. Included below are detailed assumption notes and a forecast per country 365 days ahead (starting with the day of the first case).
Why forecast using pre-lockdown data as input?
Allows for a more accurate estimate of the eventual number of official cases, actual infections and fatalities; since the data is not distorted by the impact of an artificial lockdown (which in turn distorts estimates of model parameters). Only data available pre-lockdown is used to fit the model.
The forecast model provided a good approximation to reality for South Africa. The 365 day forecast released on the 18th of May 2020 (for SA) forecast fatalities between 12,074 and 48,287 by the end of the 365 day window (counting from the first fatality in March 2020 and ending 6 March 2021). Actual fatalities as of 6 March 2021, equalled 50,566, just slightly higher than the upper forecast boundary. As expected, the forecast peak is far to early, as the model ignores artificial delay caused by lockdowns. However timing the peak was considered less important than estimating the eventual total number of fatalities over the forecast range.
Forecast slide decks by country
The primary impact, under the no lockdown scenario, is that the speed at which infections spread is greater, and so the peak is reached more quickly.
The forecasts below give an indication of what might have happened if voluntary social distancing had been the only non-medical intervention. The models are estimated based on actual data prior to a lockdown. The ‘official’ daily new case data was pre-processed. To improve the accuracy of the forecasts the official daily new case data was adjusted to estimate the actual number of new infections each day. This was further adjusted to account for the number who are currently infectious (contagious) using a moving window. The two days prior to the lockdown were excluded from the data used to fit models (in each country) since some may have changed behaviour in anticipation of a lockdown.
Original code from Learning Machines. Acentric adapted the original code to improve accuracy. The original code input 'cumulative cases' which is not suitable for a SIR model. A new module was created to estimate daily new 'infectious' (as opposed to 'infections' or 'cases' which are entirely different metrics) to provide the correct input and to automate data gathering and reporting.