Forecasting in uncertain times: modelling lessons from COVID-19

We have been exposed to multiple COVID-19 estimates, including the number of contagious, fatalities, rates and other metrics. These forecasts consistently change and we have seen significant differences in the projected numbers. The bombardment of estimates can become overwhelming and confusing, leaving people unsure of what to believe. The question is, why are some estimates so optimistic while others project a tragedy and why is there such large spreads in the forecasted numbers?

COVID-19 has provided an insight into the complex nature of mathematical modelling. Forecasts are estimates of future events, where the predicted result is calculated based on the value of one or multiple variables. For example, if we try to estimate the total number of COVID-19 cases in the future using known information, we could expect that:

Number of COVID-19 cases in the future = (cases today) + (new cases) – (recoveries) – (fatalities).

If all the data is available, this should be a very simple calculation. However, the reality is a bit different. When we think about the components of this simple formula independently, we realise there is more to it. The number of recoveries, for example, requires a different model on its own.

Variables became complex very quickly and we have multiple models within a model. Commonly, to simplify things, assumptions will be utilised. Including assumptions in modelling means that the forecasts will be based on beliefs accepted as true without proof. Different assumptions will produce different estimates and the decision of which assumptions to use will likely be taken based on empirical evidence.

Translating this short example into a business environment, every organisation requires forecasts to make strategic decisions, even if they are dealing with high levels of uncertainty and modelling complexity. There is no simple solution to this, however there are some considerations that could be made when predictive models (e.g. forecast revenue) are built and high levels of uncertainty exist:

  • Get the right people across the organisation to review the problem and produce a model based on an informed consensus view. 
  • Include people from different backgrounds and experience in the discussion making process, to ensure assumptions are made in a holistic manner.
  • Maintain a healthy level of scepticism about your model, it will provide an indication of the future but never the absolute truth.
  • Understand the value of ‘scenario testing’ where multiple calculations are performed using the same model but with different inputs and assumptions. It will prepare your business for different outcomes.
  • Understand the value of ‘stress testing’ and use inputs and assumptions that produce extreme scenarios. Worst case scenarios have usually a low probability of occurrence. If you expect the worst, any other scenario will represent a win and therefore reducing disruption within an organisation.