Agile software engineering practices have become the standard work management tool for modern software development teams. Are these techniques applicable to analytics, or is the nature of research prohibitively distinct from the nature of engineering?

In this post I discuss some adjustments to the scrum methodology to make the process work better for Analytics and Data Science teams.

If you have not already read them, you can find

When managing a team of Analysts and Data Scientists, we want to leverage all of the best things about the scrum methodology while adjusting for the things that don’t work so well. The general theme of this post is how to carve out some room in scrum for exploration, research, and iterative problem definition.

Key Adjustments to Scrum for Analytics

Time-bound spikes for research

In order to address the difficulties in using scrum for research and exploratory data analysis, I recommend leveraging the notion of a time-bound “spike” on a particular subject. Here are a few examples of spikes you could pursue:

All of these are interesting technical questions that require a Data Scientist or Analyst to answer, but 1) it is unclear exactly what the right path to completing the task is, and 2) you can imagine spending nearly infinite amounts of time on each of these questions if you let yourself bury into the rabbit-hole.

To address these issues, you should treat these as time-bounded tasks. For example, you might say that the deliverable is a one-page document outlining the results of the investigation, and you should not spend more than two working days on the task (or whatever is appropriate).

If the task seems like it is going to take longer, that is a good time for the Analyst or Data Scientist to come up for air and discuss what they have found with the group before pursuing the task any further.

Build in slack time for exploration

Google is famed for having a “20%” time (though it is unclear to what extent it is real) that engineers can use to work on projects outside of their normal jobs that they believe will benefit Google.

For Analytics teams, building in some slack time to focus on research and exploration is critical to identifying the next big opportunity for the business. If all of your Analysts and Data Scientists are fully allocated against to-do list checking and reporting concerns, you may be missing huge opportunities to drive impact.

Now, I don’t necessarily think it is a good idea to allocate one fifth of someone’s working time to goofing off. It is important to implement a program like this with accountability and oversight. Here are some ideas for how to make something like this work:

I like to think of this time as R&D investment time: the exploratory tasks of this quarter have the potential to become the headline projects of next quarter if we are able to identify an exciting area of opportunity. Of course, the nature of R&D work means that not every project will work out – pharmaceutical companies may test thousands of potential treatments before finding one that can make it through clinical trials – similarly, this sort of research is an investment with risk (how to prioritize work in such an environment is a subject for another blog post).

Acceptance Criteria to “write the next story”

Often in data science and analytics research tasks, it is difficult to map out the course of the project from day 1. We expect (and want!) the things we learn in the early parts of the research phase to shape the future phases. While this can make long-term planning difficult, it is important to build the right amount of flexibility into your process.

One trick that I like is to have the person doing the task write the story for the next task. Let’s take our example spike story from above “Research the state-of-the-art in LTV modeling for e-commerce companies”. The ACs for this story could be:

By chaining stories together this way, we can continue using our scrum methodology while still allowing the individual doing the research the flexibility to iterate on their work plan as the work progresses. By actually turning each step in the project into a well defined story, we can inject both peer-review and accountability into the sometimes opaque research process.

Peer-review in show-and-tell

At the end of the sprint, we like to bring the team together to “demo” what they have built. In software engineering scrum practice, we expect the engineers to demo “working software.” However, for analytics and data science we may do some work that involves building working software but there is lots of work (e.g., research) that does not necessarily involve anything demoable.

In addition to doing “software” demos, it can work well to treat the time similar to how research presentations work in academia. Each scientist can share their research and the team can use the time to ask clarifying questions, poke holes in the methodology, and suggest areas for future research. This time then serves a bunch of purposes:

Conclusion

Scrum is a great tool for project managing analytics and data science work. In this post I have suggested a few adjustments I like to make to the process to make it more amenable to doing the sort of research work that Data Analysts and Data Scientists spend their time on.

For me, the key thing to remember is that the most important thing is generating valuable work, and whatever process you use should be subordinate to that goal. To the extent the process helps you generate valuable work, you should use it, and to the extent that it distracts you from that prime directive, you should adjust it!