Quicksilver LB
Follow Us



Why Data Science Fails

Eye of the Storm: Natural disasters, the insurtech market, and other musings from Bob Frady, CEO of HazardHub


Throughout our career, the principals at HazardHub have been involved in a number of data science experiments and tests. We’ve learned a lot about what makes for a successful – and unsuccessful – data science engagement.

Here’s a clue: Data science almost never fails. It’s math and code, most of which works really well.

The “failure” in data science is not actually in the data – or the science – at all. But data science projects that “fail” all have one element in common.

In the words of Led Zeppelin, it’s a “Communication Breakdown” that’s primarily the fault of the business, where the business does a poor job of defining what they’re looking to accomplish.

That’s right. It’s almost always the fault of the business and its failure to set down a good plan.

While data science teams aren’t totally blameless, either, almost always, failed data science efforts come down to poor strategy from the business.

One of the big obstacles in harnessing the power of data science is that it involves a lot of concepts that are difficult to communicate between the business and the data science teams. Concepts that excite data scientists can also completely terrify most business users. Mention things like “extreme gradient boosting” or “random forest” to a business user, and you’ll get stares as blank as an unpainted canvas. Conversely, if you mention “explainability” and “simplicity” – terms that motivate the business user -- and the data scientists can sometimes check out because you’re messing with the purity of their statistical approach.

For the business user, data science is not like a refrigerator that you open and say, “I don’t know what I want, but I’ll know it when I see it.” You should have a coherent, explainable, and well thought out set of idea about what you’re trying to accomplish. You should also be willing to adapt your views based on what the data science teams turn up.

For data scientists, purity of the statistical output is a terrible measure of success. Sometimes, close enough is good enough if “close enough” can be explained and digested by the business. There are many, many statistical choices you can make to get to a desired outcome. But if it doesn’t meet a business objective, people will talk badly about you behind your back. Which is never fun.

So, here’s a simple process to make sure your next data science project does not “fail.”

  1. Have the business write up a strategic goal about what they are trying to accomplish – including what would constitute a “success” for the business. (Hint: “See if there’s anything there” is not a strategic goal. It’s a hope.) Make sure the goals are easy to understand. There are a lot of data scientists who have no real idea what the business implication is for your request. Use this opportunity to explain it to them.
  2. Meet with your data science team to fully explain what you are attempting to accomplish. If you’re on the data science team, now is the time to ask questions. If the data science team does not ask any questions – or the business cannot pinpoint what they’re looking for – it’s time to go back to Step 1. The goal in this step is not for the business to “sell” an outcome, but instead to have an open discussion about how both teams can drive to success.
  3. Constantly communicate interim results. This critical step often gets overlooked, usually because the business falls down on keeping everyone accountable (after all, data science is scary), or the data science team holds its cards too close (because they’re afraid of changing requirements). If you’re not checking in at least once a week, you’re not checking in enough. Make sure that you leave sufficient time to dig into the preliminary results as it can turn up issues that can often lead to a better outcome. While that means that the requirements may change, it’s actually a good thing.

Data science can be a powerful and valuable tool to help win in the marketplace. Follow these steps to help unlock the power of your data science teams.






Featured articles

Guidewire MR



The Email Chat is a regular feature of the ITA Pro magazine and website. We send a series of questions to an insurance IT leader in search of thought-provoking responses on important issues facing the insurance industry.


April 5th – 7th, 2020
The Diplomat Resort
Hollywood, FL
Become a member today to receive updates – www.itapro.org/MR


only online

Only Online Archive

ITA Pro Buyers' Guide

Vendor Views

Partner News