How to Handle 2020 in Your Predictive Models

Data Science

For many who work in predictive analytics, historical data is their bread and butter. In 2020, that bread has gone stale and the butter spoiled, as a variety of stress scenarios have uprooted patterns completely.

People are spending less. So much so, that iconic companies from Gold’s Gym to JCPenney have filed for bankruptcy. Other areas, like WFH tools, are seeing unbelievable growth.

Years ago, analysts predicted that retirees would open up 55 million jobs by 2020. The situation turned 180 degrees, as 46 million people filed for unemployment this year.

No matter where you look, long-term historical data is no longer a solid predictor in 2020.

Capture Local Behavior Instead

However, this doesn’t mean that you should throw predictive analytics out the window — quite the contrary.

Of course, models that rely on historical patterns will break when there are shocks in the system. That means we need to build models that rely more on short-term, local data, which captures new behavior.

For example, cruises are typically in high season in August, so should a cruiseline (some are open since June 1) use historical data from August 2019 to predict sales in August 2020? Obviously not. But they could try to predict sales based on the number of customers in the past 6 weeks and the type of tickets and amenities they bought.

The same principles apply to a bakery, an e-commerce store, or a SaaS app.

Use-case Examples

For many predictive analytics use-cases, “time” may not even be a meaningful feature. For instance, if you wanted to predict employee attrition, you might use features like job satisfaction, performance report ratings, and how many days they take off.

To give another example, a SaaS company might be interested in analyzing churn to increase CLV, using features like the customer’s tenure, what product plan they’re on, their age, and their frequency of usage.

You only need to take a cursory look at some public datasets to realize that there are countless features to use in predictive models that aren’t date or time.

Reporting

Finally, the whole goal of data analytics is to gain actionable insights from data, which is only possible once decision-makers in the organization understand the analysis.

Given all the craziness going on, there should now be a caveat when presenting any predictive analytics in an industry that’s affected: These results are uncertain.

Predictive analytics is just that, predictive, so it’s never certain, but stakeholders need to understand that now is a particularly hard time to predict the future.

Conclusion

A variety of economic shocks have thrown a wrench in the usefulness of long-term historical data, but you can still take advantage of the power of predictive analytics by focusing on local, shorter-term data, as well as features besides “time.”

You can gain predictive insights for free on Apteo by creating a workspace and uploading your data in just a few minutes.

---

Article photo by Zach Vessels on Unsplash

Frederik Bussler

Frederik Bussler is the Founder of the Security Token Alliance. As a public speaker, he has presented for audiences including IBM, Nikkei, Slush Tokyo, and the Chinese government, and is featured in outlets including Forbes, Yahoo, Thrive Global, Hacker Noon, European Commission sites, and more. Recently, he represented the Alliance as a V20 delegate.