Data Science during Covid-19 - All of your models are broken

?The current pandemic represents a paradigm shift for data science and analytics. Whatever business you're in, your customers have been affected. B2B, retail, consultancy, construction; all customers have been affected. Modern data science techniques have never had to deal with such a seismic customer behavioural shift. If you have a predictive model that relies on past customer behaviour, that model won't work anymore!

Customer analytics is all about understanding customer behaviours using only data and the skill of a data scientist. Its mostly concerned with using the past to predict the future. To understand, using collected data, any trends that might exist or insight into your customers. It is unlikely to rely directly on a customer's behaviour to predict that same customer's future intent. More likely, it tracks groups of customers who behave in a particular way. Think, "people like you, bought this product". 

This insight is used extensively by sales and marketing teams to target customers. Similar insights are used to define debt collection strategies, segment a company's customer base or even identify target customers for a sales campaign. 

Its common that some customers personal circumstances will change over time. People get married, have kids, buy houses and lose jobs, but normally, not all customers will see their personal circumstances change at the same time. This fact means that analytics models are pretty robust. You may have heard about precision and recall or model confidence, in fact all models rely on the fact that past actions will predict future behaviour. 

At this time, data science teams should be re-evaluating every customer related predictive model. Management should be asking for the latest model performance indicators so that things return to normal they can compare performance. Notably, it may be up to astute and experienced marketing managers to provide insight into customer sentiment and the economic outlook long before predictive modelling can gain its foothold again.