• Share
  • Share
Man pointing to printed graphs and charts

Edging forward with advanced analytics

5 strategies to rock out with advanced analytics—and change the way you do business.

Sabah Sadiq | December 3, 2015

Advanced analytics is more than just fancy math.

It’s a change in the way you do business. It’s a collaboration between data people and decision people—a collaboration that requires everything from data science to strategy and culture transformation expertise.

At Slalom, we know how to implement analytics to affect change. It requires an end-to-end solution with models that are simple to understand, easy to execute, and agreed upon by everyone.

We use these five strategies in our advanced analytics projects to give us an edge in successfully driving analytical change throughout an organization—and help our clients move the needle.

1. Understand that model buy-in is the key to affecting change

Models are awesome.

They provide insight into what’s driving human behavior. They unearth actions that can lead to positive outcomes. They help identify where your efforts should be concentrated to maximize value. However, if you can’t get both business leaders and end users to buy into your predictive model, all of those possibilities are lost.

According to the Fortune Knowledge Group Fund, 61% of business leaders reported that they trust real-world insight over analytical models when making decisions; 62% stated that they tend to trust their gut.

While their gut just might be right, you need to know to know why it’s telling them something and then try to capture that in the model so it makes sense to them.

This is why data scientists must work closely with key stakeholders to get buy-in for data sources, key predictive drivers, and the modeling algorithm itself. Without this collaboration, either—or worse, both—the business decision makers or end users may resist the solution.

2. Simplicity is best

A complex algorithm may sound cool, but if it’s infinitely harder to implement, it’s probably not worth modeling. An external variable may marginally increase your accuracy but could be expensive to buy. This doesn’t mean that you can’t use more sophisticated models (or something similar)—most algorithms can easily be coded in a structured language—but take care to choose algorithms and data that address the business problem at hand and are easily executed from a technical standpoint.

Take the case of Netflix. In 2009, Netflix hosted a public competition to predict user film ratings. After months of intense competition, a highly accurate solution was chosen as the winner—yet never implemented. Although complex algorithms can improve accuracy in predictive models, they can be difficult to implement.

A simpler model is a better model.

3. Understand the difference between modeling and scoring

According to Eric Schmidt, former CEO of Google, every two days we create as much information as we did from the beginning of time until 2003. That’s a lot of data.

Once a data scientist builds a predictive model, it must be applied to the data—in other words, the data needs to be “scored.” Scoring involves tagging each data point with the output from the mathematical model: a common example would be assigning each customer into a customer segment. In order for the model to be as real-time as possible, it must be scored based on the most recent and accurate information available.

There are two ways to score data: you can rerun the model every time you need a new answer, or you can translate the model into a structured language like SQL. Thanks to modeling packages in Big Data systems, the former is possible, but not without its issues. A small change in the underlying data can cause unstable predictions; there’s less transparency around the model formula; and processing time can increase significantly. For now, when it comes to quickly and consistently scoring large data sources, SQL reigns king.

4. Test the model constantly and vigorously

Predictive models often replace a longstanding way of doing business. Whether an upgrade to a basic forecast or a replacement for fly-by-the-seat-of-your-pants guessing, predictive models represent change.

In order to validate that the change is worthy, test models from both a statistical and common sense perspective. Vigorous use-case testing allows IT to verify that the models are accurate, and the business to gain familiarity and comfort with the model. Accuracy and adoption are key.

5. Develop visual tools to allow users to interface with the model

Ultimately, models are created to help people make decisions. To do so, people need to be able to touch and feel the model on a daily basis. Interactive dashboards are essential to truly give the data back to the people.

Best-in-breed analytical implementations recognize this and are becoming more tablet-friendly. With a few clicks or taps, users can buy optimal advertising spots to target specific customers, generate the most effective media campaigns, and take steps to reduce churn of their most valuable employees. All this can be done with a few clicks, pretty graphs, and rock-solid analytics.

So, what’s the difference between building a predictive model and successfully implementing one? At Slalom, we think it’s a lot.

Successful model adoption requires a deep understanding of business. Only when data science collides with strong cross-functional teams and implementation skills will advanced analytics truly sing.

Sabah Sadiq is no longer with Slalom.


Start a conversation

What’s on your mind? Let’s explore the possibilities.