forecasting employee turnover: how to incorporate impact of coronavirus I am working on a model for my company to forecast employee turnover (employees leaving the company). My model has worked well in the past and been fairly accurate in its predictions. However, I am concerned that with the coronavirus situation, it will not be as accurate in future predictions. I am not sure how to adjust the model as the coronavirus situation is unprecedented. There is already a variable in the model for the BLS unemployment rate. Any ideas for what variables to include in the model to get at the impact of coronavirus, and how to include them? Any thoughts welcome.
 A: This corresponds to the fundamental assumption of supervised machine learning that we can use historical data to make predictions on new data. For time series modeling, you should be aware of the historical time period that you hope to learn from. Since you don't have data to learn from a pandemic, any insights learned from previous periods should be taken with skepticism. Even if your company existed during the Spanish Flu Pandemic, so much about society was different that the data could be useless.
The main approach I've seen is that if you have data from a previous crisis, you can use them as a benchmark. For example, you can forecast that you'll have 2x the turnover or half the turnover from the Great Recession.
In general, linear models are better at extrapolating when they encounter extreme values. Someone might love her Facebook Prophet model, but it would adapt more slowly when encountering shocks. Also be careful in looking too far in the past for feature derivation for things like lags.
The most important path you can take is approach this less from a machine learning perspective and more from a sociological perspective. Use your subject matter expertise to determine whether employees will leave. Basic exploratory data analysis matters most now.
If you want to include Covid-19 related variables, the number of deaths per capita is probably a better metric to look at than positive cases. The benefit of this is that there is a difference between 0 deaths and many deaths, but the effect in an early country like Iran or Italy is different from the effect in a country that got infected later. How you think virus prevalence will affect your company is up to your judgement.
Finally, be up front with uncertainty. Many public companies stopped sharing forward guidance during the pandemic because they have no idea what the future has in store. When this is all over, you won't want to use historical data from a pre-vaccine world.
